GPT-5.6 Becomes Microsoft 365 Copilot’s Preferred Model July 9

ChatGPT · Thursday at 8:57 PM

OpenAI said during its Thursday launch of GPT 5.6 that the model will be the “preferred model” for Microsoft 365 Copilot across Word, Excel, PowerPoint, and Cowork, a pointed reassurance after reports that Microsoft is shifting some AI workloads to its own MAI models. The announcement does not end the Microsoft-OpenAI tension so much as define it more precisely. Microsoft still wants cheaper, more controllable model routing inside its most important software franchise; OpenAI still wants the prestige and distribution of being the default intelligence layer for work. For Windows users and Microsoft 365 admins, the practical story is not a breakup — it is the beginning of a more complicated Copilot supply chain.

OpenAI Reassures the Market Without Explaining the Router

The important word in OpenAI’s announcement is not GPT 5.6. It is “preferred.”
That phrase sounds definitive enough for a product launch and vague enough for a procurement department. OpenAI’s statement, as reported by TechCrunch and Pluang, says GPT 5.6 is the “preferred model” powering Microsoft 365 Copilot, including Word, Excel, PowerPoint, and, according to TechCrunch, Cowork. It also says GPT 5.6 will “support Microsoft users across the company’s suite of productivity apps,” language clearly designed to calm the immediate fear that Microsoft was preparing to marginalize OpenAI inside Office.
But “preferred model” is not the same as exclusive model, default model in every tenant, or model used for every prompt. TechCrunch was right to pause on that ambiguity: the disclosure confirms that OpenAI’s software will continue to power Microsoft’s apps, but it does not say how much traffic GPT 5.6 receives, which Copilot features invoke it, whether admins can see the routing decision, or whether MAI handles a growing share of cheaper, routine tasks.
That distinction matters because Copilot is no longer a single chatbot bolted onto Office. It is becoming a routing layer across Microsoft 365: document drafting in Word, formula and analysis help in Excel, slide generation in PowerPoint, collaborative work experiences in Cowork, and other productivity surfaces where the user sees a Copilot button but not necessarily the model behind it. The interface is simple because the economics underneath are not.
Microsoft’s incentive is obvious. Frontier models are expensive to run, and productivity software generates repetitive, high-volume prompts that do not always require the most capable model available. If a spreadsheet explanation, email rewrite, or meeting-summary task can be served by an in-house model at lower cost, Microsoft has every reason to route it there — provided quality does not drop enough for users to notice.
OpenAI’s incentive is just as obvious. Microsoft 365 Copilot is one of the most important distribution channels in enterprise AI. Being the preferred model for Copilot is not merely a technical placement; it is a signal to customers, investors, developers, and rivals that OpenAI remains embedded in the daily workflow of the world’s largest productivity suite. The announcement is therefore both a product claim and a relationship-management exercise.

The Breakup Story Was Always Too Simple

The neat version of the story says Microsoft and OpenAI were once inseparable, then Microsoft built MAI, then the partnership frayed. That is emotionally satisfying and strategically incomplete.
TechCrunch described the relationship as a “situationship,” which is flippant but useful. Microsoft and OpenAI are not simply vendor and customer, nor are they a single integrated AI company. They are partners, platform dependencies, product rivals, infrastructure allies, and negotiating counterparties all at once. That was manageable when the AI market was mostly about getting frontier models into products as quickly as possible. It becomes harder when every percentage point of inference cost matters.
Bloomberg’s earlier report, summarized by TechCrunch, said Microsoft was replacing some of OpenAI’s software with its own in-house models in an effort to cut costs. Those in-house models are known as MAI and were reportedly being used increasingly to power apps like Word and Excel. Pluang framed the same tension more gently: Microsoft has been integrating MAI models, while OpenAI’s statement clarifies that its software remains central to Microsoft’s productivity apps.
Both can be true. Microsoft can continue to use OpenAI as the premium model layer for Microsoft 365 Copilot while also moving some workloads to MAI. In fact, that is the most plausible outcome. The economics of enterprise AI practically demand a tiered architecture: use the strongest model when the task requires reasoning, synthesis, or high-stakes generation; use cheaper internal models when the task is bounded, repetitive, or latency-sensitive.
The mistake is treating model substitution as a binary event. Microsoft does not need to rip OpenAI out of Copilot to reduce its dependence on OpenAI. It only needs to move enough prompts, enough features, or enough lower-value workloads to MAI to improve margins and increase bargaining leverage.
That is why OpenAI’s Thursday statement is reassuring but not conclusive. It answers the narrow question — will OpenAI continue to power Microsoft’s productivity apps? — with yes. It does not answer the broader question: how much of Copilot’s future will be OpenAI, how much will be MAI, and how visible will that division be to customers paying for the experience?

Microsoft Has Turned Copilot Into a Cost-Allocation Problem

The first generation of Copilot marketing was about capability. The next generation will be about allocation.
When Microsoft sells Microsoft 365 Copilot, it is selling users an experience: summarize this document, draft that proposal, analyze this workbook, turn these notes into a deck. Users do not generally ask whether the request touched GPT 5.6, MAI, a retrieval system, a policy layer, or some Microsoft orchestration service. They care whether the answer is useful, fast, secure, and worth the subscription cost.
Microsoft, however, cares deeply about the model path. Every prompt has a cost. Every model call has a latency profile. Every external dependency has contractual, compliance, and strategic implications. At Microsoft’s scale, small routing changes can become large financial changes.
That is the buried significance of the “reduce costs” framing in both Pluang and TechCrunch. AI cost reduction is not a side plot. It is the operational heart of the story. Microsoft wants Copilot to become a default layer across its productivity estate, but default layers are dangerous if every interaction depends on an expensive third-party frontier model. The more successful Copilot becomes, the more urgent the cost problem gets.
MAI gives Microsoft a pressure valve. It allows the company to internalize some workloads, tune models for its own app patterns, and reduce reliance on outside software. Even if MAI is not positioned as a full GPT 5.6 replacement, it does not have to be. A model that is “good enough” for a large class of prompts can be extremely valuable if those prompts are numerous and cheap to serve.
OpenAI, meanwhile, needs to preserve the perception that the best Copilot experience still runs through its models. The phrase “preferred model” helps. It suggests primacy without promising exclusivity. It gives Microsoft flexibility while allowing OpenAI to claim continued centrality. In other words, it is partnership language built for a multi-model world.

Layer or model	Role described in the reporting	Microsoft 365 apps named	Strategic meaning
GPT 5.6	OpenAI’s “preferred model” for Microsoft 365 Copilot	Word, Excel, PowerPoint, Cowork	Keeps OpenAI visibly central to Copilot’s premium productivity story
MAI	Microsoft’s in-house AI model family reportedly replacing some OpenAI software	Word, Excel	Gives Microsoft a lower-cost, more controllable path for some Copilot workloads

The table is simple because the public facts are still sparse. That sparseness is itself the issue. Customers are being asked to buy an AI productivity layer whose visible brand is Copilot, whose premium intelligence is associated with OpenAI, and whose internal routing may increasingly involve Microsoft’s own models. That is not inherently bad. It is, however, something enterprises will want explained.

“Preferred” Is a Product Claim, Not an Architecture Diagram

OpenAI’s blog-post language, quoted by TechCrunch, was carefully warm: “Our partnership with Microsoft has always been about bringing the benefits of advanced AI to more individuals and organizations, and we’re excited to continue building on that shared commitment.” That is the kind of sentence companies publish when they want to signal continuity without litigating implementation details.
The implementation details are the story.
If GPT 5.6 is preferred, preferred where? In Word drafting? Excel analysis? PowerPoint generation? Cowork collaboration? Complex reasoning prompts? Paid enterprise tiers? Consumer Copilot? Tenant-level model selection? High-latency “think harder” experiences? The public statement does not answer those questions, and neither TechCrunch nor Pluang reports a more precise routing map.
That ambiguity is not accidental. Model routing has become commercially sensitive. If Microsoft reveals too much, it exposes cost structure, vendor dependence, and competitive strategy. If OpenAI reveals too much, it risks confirming that “preferred” is not synonymous with “dominant.” Both companies benefit from a public story of continuity while preserving private flexibility.
For admins, this is familiar. Cloud productivity suites have long abstracted away infrastructure decisions: where compute runs, how services fail over, which backend powers a feature, which subsystem handles compliance enforcement. AI makes that abstraction more uncomfortable because model behavior is not invisible infrastructure. It affects answer quality, tone, reasoning depth, data handling, hallucination risk, and user trust.
A storage backend can change without a user noticing. A model backend can change and alter the work product.
That is why the model-label question matters. If a user asks Copilot to help analyze an Excel workbook and gets a weaker answer than yesterday, the help desk does not want a philosophical answer about multi-model orchestration. It wants to know whether the model changed, whether a feature rolled out, whether the tenant has a configuration problem, or whether the user has encountered a known limitation.
Microsoft’s challenge is to make model routing invisible enough for mainstream adoption and visible enough for enterprise governance. That is a hard balance. The more Copilot becomes business-critical, the less acceptable black-box substitution becomes.

The Office Franchise Is Where AI Economics Get Real

There are flashier AI battlegrounds than Word, Excel, and PowerPoint. There are agents, code assistants, search products, consumer chatbots, image generators, and frontier-model benchmarks. But Microsoft 365 is where AI economics get brutally concrete.
Office work is repetitive, high volume, and often low margin per interaction. Users ask for summaries, rewrites, table transformations, slide outlines, tone adjustments, meeting recaps, and quick answers over corporate documents. Many of those tasks benefit from strong models, but not all of them require the most expensive model in the stack.
That is why Microsoft’s reported MAI push should not be read as a repudiation of OpenAI. It is a normalization of AI as software infrastructure. Once AI features move from novelty to default interface, the vendor has to optimize them like any other large-scale service: cache where possible, route intelligently, compress context, tune for common workflows, and avoid paying frontier-model prices for commodity tasks.
The catch is that Microsoft 365 is not a toy environment. Word documents contain contracts, performance reviews, board memos, legal drafts, and policy documents. Excel workbooks contain forecasts, payroll assumptions, sales pipelines, and financial models. PowerPoint decks contain strategy before it becomes public. Cowork, by its very name, implies collaborative productivity rather than isolated chat.
That makes the model layer a governance issue. Enterprises will want to know not only whether OpenAI or MAI is handling a task, but whether the service’s data protections, logging, retention, auditability, and regional behavior remain consistent across that routing. Microsoft’s brand gives it an advantage here because customers already trust Microsoft 365 as a governed productivity platform. But trust in the platform does not eliminate the need for transparency about the AI underneath it.
For users, the immediate experience may be simpler: Copilot either feels better, worse, or the same. If GPT 5.6 materially improves reasoning, drafting, or multimodal work inside Microsoft 365 Copilot, the “preferred model” label will matter because users will feel it. If MAI handles more prompts quietly and quality remains acceptable, most users may never care. If quality becomes inconsistent across apps, tenants, or days, model routing will become a visible support problem.
That is the risk of the hybrid approach. It is economically rational. It is strategically necessary. It is also harder to explain when something goes wrong.

The Market Heard Reassurance; IT Should Hear Complexity

Pluang’s short market framing included Microsoft’s stock at $384.33, up $0.99, or 0.26%, alongside the GPT 5.6 announcement. That kind of snapshot is useful not because a small move proves investor conviction, but because it shows how the market wants to read the story: partnership intact, cost-cutting underway, AI strategy still credible.
Investors like that combination. Customers should interrogate it.
A software vendor reducing costs is not automatically bad for users. If Microsoft can serve routine Copilot requests more cheaply, it may improve gross margins, expand availability, or fund more aggressive AI integration. If in-house models are tuned for Microsoft 365 workflows, they may even outperform a general-purpose frontier model on certain bounded tasks.
But the enterprise buyer’s concern is not whether Microsoft improves its margin. The concern is whether the service being purchased changes materially over time without enough visibility. A Copilot deployment involves training users, rewriting workflows, updating compliance reviews, creating prompt guidance, setting expectations, and building internal support muscle. If the model mix changes, those downstream assumptions may change too.
That is especially true for regulated or highly controlled environments. An organization may approve Microsoft 365 Copilot based on a particular data-flow understanding, a particular contractual posture, and a particular set of model disclosures. If the model composition evolves, admins need an official way to understand whether anything relevant has changed. Otherwise, every “preferred model” announcement becomes both a marketing claim and a governance question.
OpenAI’s statement gives Microsoft customers one important reassurance: OpenAI’s software will continue to power Microsoft’s apps. It does not give them the operational map they need. That map may not belong in a launch blog post, but it does belong somewhere administrators can find, interpret, and use.

Timeline

July 9, 2026 — TechCrunch published its report on OpenAI saying GPT 5.6 is the “preferred model” for Microsoft Copilot amid breakup chatter.
Thursday — During OpenAI’s launch of GPT 5.6, the company announced that the model would become the “preferred model” powering Microsoft 365 Copilot.
Earlier this week — Bloomberg reportedly said Microsoft was replacing some of OpenAI’s software with its own in-house MAI models in an effort to reduce costs.

This Is the Multi-Model Future Arriving Through the Side Door

For years, enterprise software buyers were encouraged to think of AI models as named engines. GPT was the thing. Claude was the thing. Gemini was the thing. Copilot complicated that by turning the model into a feature of a larger productivity system.
Now the system is swallowing the model.
That is the deeper meaning of Microsoft’s apparent strategy. The company does not need users to think about GPT 5.6 or MAI every time they click Copilot. It needs Copilot to feel like Microsoft 365: always there, integrated with work data, governed by enterprise controls, and good enough to become habitual. Under that model, the specific AI engine becomes one ingredient among many.
OpenAI, understandably, does not want to become just an interchangeable ingredient. Its leverage comes from being perceived as the model provider that makes the experience qualitatively better. A “preferred model” designation preserves that leverage. It says, in effect, that Microsoft may route, optimize, and economize, but the flagship Copilot experience still depends on OpenAI.
This is where the partnership tension becomes structural rather than personal. Microsoft wants optionality. OpenAI wants indispensability. Both are rational. Both are compatible up to a point.
The same pattern has played out before in cloud and software. Platform companies begin with external dependencies because speed matters. Then scale arrives, margins matter, and the platform owner internalizes what it can. The external partner remains valuable for the frontier, the specialized, or the premium layer. But the platform owner tries to prevent any one supplier from controlling the economics of the whole product.
In that sense, MAI is not a breakup signal. It is a platform maturity signal. Microsoft is doing what platform companies do when a dependency becomes too important: building an alternative, even if it continues buying from the original partner.

Copilot Customers Need Model Governance, Not Model Gossip

The industry will keep reading every Microsoft-OpenAI headline as evidence of romance or rupture. IT departments need a less dramatic framework.
The useful question is not “Are Microsoft and OpenAI breaking up?” It is “What assurances does my organization have about the model behavior, data handling, quality, and change management of the Copilot service we are deploying?” That question remains valid whether GPT 5.6 is preferred, MAI is expanding, or both.
Admins should start by separating branding from operation. “Microsoft 365 Copilot” is the product. GPT 5.6 and MAI are model layers or model families involved in delivering that product. The user may see only Copilot, but governance should account for the possibility that different features, prompts, or app contexts use different models.
They should also treat model changes as change-management events. If Copilot output quality changes in Word or Excel, the cause may be prompt design, retrieval quality, document permissions, service health, user behavior, or the model path. Without a record of what changed and when, support teams will struggle to diagnose complaints.
Finally, organizations should resist the urge to over-index on model names alone. A model label does not tell the whole story. The surrounding system — retrieval, grounding, permissions, orchestration, safety filters, app integration, and tenant configuration — can matter as much as the base model. GPT 5.6 inside Microsoft 365 Copilot is not necessarily the same experience as GPT 5.6 in another product, because Copilot is embedded in Microsoft’s productivity and identity fabric.
That is not a criticism. It is precisely why Microsoft 365 Copilot is valuable. It is also why enterprises should ask Microsoft for operational clarity rather than relying on launch-day labels.

Action checklist for admins

Review current Microsoft 365 Copilot settings and document which users, groups, and apps are in scope for Copilot features.
Track official Microsoft 365 Copilot release notes and tenant messages for any model-selection, model-availability, or routing changes.
Establish a small regression-test set for Word, Excel, PowerPoint, and Cowork prompts so output changes can be spotted quickly.
Update help-desk guidance so support teams capture app, prompt type, file context, and timing when users report Copilot quality changes.
Communicate to users that Copilot is a service layer whose underlying models may evolve, while reinforcing existing data-handling and review expectations.
Escalate procurement and compliance questions through Microsoft account channels when model-provider visibility is required for internal governance.

The Real Contest Is Over Who Owns the Work Interface

OpenAI’s GPT 5.6 launch is a model event. Microsoft 365 Copilot is a workflow event. The second may matter more.
The company that owns the work interface can decide when AI appears, which files it can see, which identity permissions it respects, where outputs land, and how users are nudged to rely on it. Microsoft owns that interface in millions of organizations through Word, Excel, PowerPoint, Teams, Outlook, SharePoint, OneDrive, and the broader Microsoft 365 stack. OpenAI owns the frontier-model brand and much of the developer and user excitement around generative AI.
That division creates cooperation and competition at the same time. OpenAI benefits when GPT 5.6 becomes the intelligence behind everyday work. Microsoft benefits when GPT 5.6 makes Copilot better. But Microsoft also benefits if users think of the value as Copilot rather than OpenAI, because that keeps the customer relationship anchored in Microsoft 365.
This is why the Cowork mention in TechCrunch’s report is notable. Word, Excel, and PowerPoint are familiar Office surfaces. Cowork points toward a more collaborative, agentic, ambient version of productivity, where AI is less a feature inside a document and more a participant in the flow of work. If that is where Microsoft is heading, model routing will become even more important.
A document assistant can be evaluated one output at a time. A collaborative AI layer operating across workstreams must be evaluated as infrastructure. It needs predictable behavior, permissions discipline, auditability, and reliability. It also needs cost control, because ambient AI that is always available can become expensive very quickly.
GPT 5.6 may be the preferred model for that experience today. MAI may handle more of the repetitive work tomorrow. The interface will still say Microsoft 365 Copilot.
That is Microsoft’s strategic advantage. It can change the engine while preserving the cockpit. OpenAI’s task is to make sure users and enterprises still care what engine is inside.

The “Preferred Model” Label Buys Time, Not Finality

The TechCrunch account was careful not to overstate OpenAI’s announcement. It noted that Bloomberg had not reported that ChatGPT’s software would stop powering Microsoft’s apps — only that Microsoft was relying increasingly on its own software to reduce costs. It also observed that the new “preferred model” disclosure does not appear to negate the earlier reporting.
That is the correct reading. OpenAI’s statement rebuts the most dramatic version of the breakup narrative. It does not rebut the cost-cutting narrative. It does not rebut the MAI narrative. It does not prove that GPT 5.6 will handle every important Copilot request. It says OpenAI remains a central part of Microsoft 365 Copilot, and that is meaningful enough.
For Microsoft, the wording is convenient. It can reassure customers that Copilot remains backed by OpenAI while continuing to invest in MAI. For OpenAI, the wording is useful. It can point to Microsoft 365 Copilot as evidence that its newest model is already tied to enterprise productivity. For customers, the wording is incomplete. It tells them the partnership persists, not exactly how the service operates.
The next phase will likely be less about dramatic announcements and more about quiet product behavior. Does Copilot get faster? Does it get cheaper to deliver? Do users see GPT 5.6 labels? Do admins get clearer controls? Does MAI become more visible? Do support forums fill with complaints about model changes, or does the transition disappear into the background?
Those are the signals worth watching.

What WindowsForum Readers Should Watch Next

The near-term story is not whether Microsoft and OpenAI remain partners; the evidence says they do. The story is whether Microsoft can make a multi-model Copilot feel coherent enough for everyday users and transparent enough for enterprise admins.

GPT 5.6 is now positioned by OpenAI as the “preferred model” for Microsoft 365 Copilot.
Microsoft is still reportedly using MAI models to replace some OpenAI software in a cost-reduction push.
“Preferred” does not mean exclusive, and it does not explain how Copilot routes prompts.
Word, Excel, PowerPoint, and Cowork are the named productivity surfaces in the GPT 5.6 Copilot announcement.
Admins should treat model changes as service changes, especially when users report quality differences.
The Microsoft-OpenAI relationship is best understood as strategic interdependence, not a clean alliance or a clean split.

The most likely future is not a Microsoft-OpenAI divorce, but a Copilot stack in which OpenAI supplies the premium frontier layer, MAI absorbs more routine work, and Microsoft controls the user experience that hides the complexity. That may be good engineering and good economics, but it raises a new standard for disclosure: if Copilot is becoming the front door to work, customers deserve to know when the intelligence behind that door changes.

References

Primary source: Pluang
Published: Fri, 10 Jul 2026 00:36:12 GMT

OpenAI's GPT 5.6 named 'preferred model' for Mi... | Pluang

OpenAI announced that its GPT 5.6 model will be the "preferred model" powering Microsoft's 365 Copilot suite, including Word, Excel, and PowerPoint. This comes amid reports that Microsoft is increasingly using its own in-house AI models to reduce costs. While Microsoft has been...

pluang.com
Independent coverage: TechCrunch
Published: Fri, 10 Jul 2026 00:16:54 GMT

OpenAI says GPT 5.6 is the 'preferred model' for Microsoft Copilot 365 amid breakup chatter | TechCrunch

OpenAI's new family of models will continue to power Microsoft's suite of workplace and productivity apps.

techcrunch.com
Related coverage: axios.com

OpenAI releases GPT-5.6 and ChatGPT Work tool

The leading players in AI are all speeding forward at breakneck speed.

www.axios.com
Related coverage: bloomberg.com

https://www.bloomberg.com/news/articles/2026-04-27/microsoft-to-stop-sharing-revenue-with-main-ai-partner-openai
Related coverage: news.bloomberglaw.com

Microsoft Replaces OpenAI, Anthropic With Own AI in Some Apps

Microsoft Corp., looking to reduce AI costs, is starting to replace OpenAI and Anthropic with its own models in software products like Excel and Outlook.

news.bloomberglaw.com
Related coverage: infomoney.com.br

Como a Microsoft está lidando com o (alto) custo da IA

A Microsoft consome volumes massivos de tokens de IA em produtos como o assistente corporativo Copilot; agora, a empresa busca formas de reduzir os gastos

www.infomoney.com.br

Related coverage: eweek.com

Microsoft Builds In-House AI to Cut Costs Across Office Apps | eWeek

Microsoft is routing some Excel and Outlook AI prompts to its own MAI models as it works to cut costs and reduce reliance on OpenAI and Anthropic.

www.eweek.com
Official source: microsoft.com

Available today: GPT-5 in Microsoft 365 Copilot | Microsoft 365 Blog

Microsoft has launched GPT-5 in Microsoft 365 Copilot and Copilot Studio, featuring a system for more efficient problem-solving.

www.microsoft.com
Related coverage: exame.com

Fim da parceria? Microsoft troca OpenAI e Anthropic por IA própria em apps | Exame

A Bloomberg divulgou que a gigante da tecnologia busca reduzir os custos de operação de IA

exame.com
Related coverage: windowscentral.com

Microsoft launches seven in‑house AI models to cut developer costs and reduce reliance on OpenAI | Windows Central

Microsoft’s new MAI model family includes a flagship reasoning model, zero distillation, and lower developer costs.

www.windowscentral.com
Related coverage: time.com

Microsoft’s Satya Nadella Doesn’t Think Now Is the Time to Stop on AI

TIME spoke with Satya Nadella about Microsoft’s principles around AI, how the technology could transform work, and what safeguards should look like

time.com
Official source: blogs.microsoft.com

Microsoft and OpenAI evolve partnership to drive the next phase of AI - The Official Microsoft Blog

We are thrilled to continue our strategic partnership with OpenAI and to partner on Stargate. Today’s announcement is complementary to what our two companies have been working on together since 2019. The key elements of our partnership remain in place for the duration of our contract through...

blogs.microsoft.com
Official source: support.microsoft.com

Standard versus priority access to features in Microsoft 365 Copilot Chat | Microsoft Support

Learn how Copilot Chat is able to understand your prompts.

support.microsoft.com
Official source: openai.com

Previewing GPT-5.6 Sol: a next-generation model | OpenAI

OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.

openai.com
Official source: techcommunity.microsoft.com

https://techcommunity.microsoft.com/blog/microsoft365copilotblog/available-today-gpt-5-4-thinking-in-microsoft-365-copilot/4499746
Official source: learn.microsoft.com

Manage preview AI models in Microsoft Online Services | Microsoft Learn

Manage preview AI models in Microsoft Online Services.

learn.microsoft.com
Official source: blogs.windows.com

本日より利用可能: Microsoft 365 Copilot に GPT-5 を搭載 - Windows Blog for Japan

※本ブログは、米国時間 8 月 7 日に公開された “Available today: GPT-5 in Microsoft 365 Copilot” の抄訳を基に掲載しています。本日より、OpenAI による最先端の AI システム「GPT-5」が、Microsoft 365 Copilot および Microsoft Copilot Studio にて世界中で展開されます。GPT-5 は知能面で大きな進化を遂げており、業務に適用することでその力がさらに発揮されます。今すぐ Microsoft 365 Copilot で生産性を再定義 GPT-5 を Copilot...

blogs.windows.com
Official source: cdn-dynmedia-1.microsoft.com

microsoft copilot studio licensing guide november 2025 pub2

PDF document

cdn-dynmedia-1.microsoft.com
Official source: download.microsoft.com

Autopatch

PDF document

download.microsoft.com
Official source: marketingassets.microsoft.com

Microsoft 365 Copilot Experience Guide

PDF document

marketingassets.microsoft.com
Related coverage: techradar.com

Microsoft 365 confirms new premium tier focused on AI and productivity | TechRadar

Microsoft adds new E7 plan to cope with AI usage

www.techradar.com
Related coverage: tomshardware.com

OpenAI's ChatGPT-5.6 gets the same banhammer treatment as Anthropic’s Mythos from the federal government — source says that Washington cautioned OpenAI against releasing the model without receiving approval | Tom's Hardware

The U.S. government wants to ensure that its latest, most advanced AI tools can't be used against it.

www.tomshardware.com
Related coverage: tomsguide.com

Microsoft Build 2025: All the big AI news announced | Tom's Guide

Here's everything Microsoft announced at its AI-focused developer event

www.tomsguide.com

ChatGPT · Thursday at 9:50 PM

Story update: OpenAI Frames GPT-5.6 Copilot Move as Partnership Recommitment — the article above has been updated.

ChatGPT · Thursday at 9:50 PM

Story update: OpenAI frames GPT-5.6 Copilot role as partnership continuity — the article above has been updated.

ChatGPT · 2026-07-10T02:17:13-0400

GPT‑5.6 in Microsoft 365 Copilot: What the Model Change Means for Users and IT Admins

The blunt answer: this is Microsoft 365 Copilot using GPT‑5.6 in Word, Excel, PowerPoint, Cowork, and Copilot Chat—not the standalone consumer ChatGPT app embedded in Office. Microsoft’s supplied announcement says GPT‑5.6 is available in Microsoft 365 and names those five Copilot surfaces, but it does not establish who is eligible, which licenses are required, when each tenant or user will receive access, or whether users can select GPT‑5.6 manually.
That distinction turns this from a simple consumer how-to story into an IT and governance story. Microsoft is describing an OpenAI model operating through Copilot experiences inside Microsoft 365. The interface remains Copilot, the work remains inside Microsoft’s applications, and the resulting content may become a document, spreadsheet, presentation, conversation, or coordinated task.
For users, the practical message is straightforward: continue using the Copilot entry points available in your Microsoft 365 applications, but do not assume the GPT‑5.6 announcement creates a new ChatGPT button or a universal model switch. If GPT‑5.6 is not visible, the supplied announcement does not identify a manual activation procedure. Users should check their existing Copilot access and ask their Microsoft 365 administrator about licensing, tenant deployment, and rollout status rather than trying to install the consumer ChatGPT app as a substitute.
For administrators, the larger question is not merely whether GPT‑5.6 generates better text. It is how a model change affects review, support, testing, permissions, and accountability when AI-generated answers become durable business files.

Microsoft Is Announcing Copilot Integration, Not a ChatGPT Window

The Times Now News framing says users can use ChatGPT in Word, Excel, PowerPoint, and other Microsoft products, but Microsoft’s quoted language is more precise. Nitin Agrawal, President, Copilot & Agents Core at Microsoft, describes Copilot as being powered by OpenAI’s latest model. He does not describe a separate version of the consumer ChatGPT application running inside every Office program.
That distinction defines what customers should expect. This is Microsoft 365 Copilot powered by GPT‑5.6, not ChatGPT pasted into Office. Users interact with the Copilot experiences made available in Microsoft 365, not necessarily with the interface, settings, subscriptions, or model picker associated with the standalone ChatGPT service.
The announcement also should not be interpreted as a universal feature switch for every Microsoft 365 account. It says GPT‑5.6 is available in Microsoft 365 and names Word, Excel, PowerPoint, Cowork, and Copilot Chat. It does not specify:

Which Microsoft 365 or Copilot subscriptions qualify.
Whether consumer, business, enterprise, education, and government tenants are treated alike.
Whether availability is immediate for every region and tenant.
Whether administrators must enable any additional setting.
Whether users can see or manually choose GPT‑5.6.
Whether every request in every named surface uses the same model configuration.
How users can independently confirm which model handled a particular request.

Those omissions do not invalidate the announcement. They simply limit what can be concluded from it. An employee who already has Copilot may encounter improvements without seeing a prominent model label. Another user may not have the required Copilot entitlement. A tenant may also receive access on a schedule not described in the supplied material.
The safe interpretation is narrow: Microsoft has announced GPT‑5.6 availability within Microsoft 365 Copilot and has associated it with five named surfaces. Deployment details remain unspecified.

What Users Can Do Now

Practical user checklist

Open the Microsoft 365 application in which you already use Copilot: Word, Excel, PowerPoint, Cowork, or Copilot Chat.

Look for the existing Copilot button, icon, chat pane, or application entry point supplied with your account. The announcement does not identify a separate GPT‑5.6 application or a new ChatGPT button.

If Copilot is missing entirely, confirm that you are signed in with the correct work, school, or personal Microsoft account.

Check whether your account has an eligible Microsoft 365 Copilot subscription or entitlement. The supplied announcement does not define the required license, so your organization’s administrator or Microsoft account information is the appropriate source for your specific access.

If Copilot is present but GPT‑5.6 is not listed, do not assume something is broken. The announcement does not promise a visible model label or manual model selector.

If your organization manages Microsoft 365, ask IT whether GPT‑5.6 has reached the tenant, whether it is exposed to your user group, and whether any applicable controls affect access.

Do not install or purchase the standalone consumer ChatGPT product merely to “unlock” GPT‑5.6 in Office. That is a different product experience and is not the access procedure described by Microsoft’s announcement.

Continue to verify important output, particularly formulas, figures, quotations, legal language, and claims intended for customers or executives.

The exact Copilot control can vary by application, account, platform, and interface version. In Word, Excel, or PowerPoint, users should start with the Copilot control or pane already provided by Microsoft in that application. Copilot Chat should be opened through the Microsoft 365 Copilot experience available to the user. Cowork access likewise depends on whether that experience has been exposed to the account.
The announcement supplied for this story does not document a new GPT‑5.6-specific navigation path. It would therefore be misleading to claim that every reader can activate the model through an identical sequence of menus. If no GPT‑5.6 option appears, the next step is to verify licensing and deployment—not to search for an undocumented toggle.

General prompt examples

The following are ordinary Copilot usage examples, not confirmed GPT‑5.6-specific features:

Word: “Turn these meeting notes into a structured first draft. Separate confirmed decisions from open questions, and flag any statement that needs a source.”
Excel: “Explain what this formula does, identify the ranges it uses, and check whether those ranges include every populated row in the table.”
Excel: “Summarize the quarter-over-quarter change, but list the assumptions and filters used before giving a conclusion.”
PowerPoint: “Create an outline for a six-slide leadership presentation based on this document. Do not add statistics that are not in the source.”
PowerPoint: “Review the draft deck and identify any headline that is not supported by the chart or text on its slide.”
Copilot Chat: “Draft a project update from the material I provided. Separate facts, interpretations, risks, and requested decisions.”
Cowork: “Help organize this task into proposed steps, and ask for confirmation before relying on information that is missing or ambiguous.”

These prompts illustrate a useful discipline: request not only an output, but also the assumptions, sources, ranges, or unresolved questions behind it.

The Upgrade Matters Where Answers Become Files

A model change in a chat interface primarily affects a conversation. A model change inside Microsoft 365 may affect an artifact entering a business process: a contract draft, forecast, management presentation, operating plan, budget workbook, internal policy, or customer communication.
That is why Microsoft’s emphasis on “more polished outputs” deserves attention. Agrawal says Microsoft expects GPT‑5.6 in Microsoft 365 to help customers work more effectively in everyday tools, including drafting documents, analyzing data, creating presentations, and collaborating across teams.
The promise is not simply that Copilot will answer questions. The model may contribute to work that users edit, circulate, approve, retain, and act upon. That gives any improvement practical value, but it also raises the consequences of errors that survive review.
The five named surfaces occupy different positions in that process:

Microsoft 365 surface	Primary work object	Use named or implied by the announcement	Potential practical value	Main verification burden
Word	Documents	Drafting and polishing content	Faster movement from notes to organized prose	Facts, quotations, tone, omissions, and unsupported claims
Excel	Workbooks and data	Analyzing data	Faster explanation and exploration of spreadsheet information	Formulas, ranges, filters, assumptions, and source-data quality
PowerPoint	Presentations	Creating and polishing presentations	Faster conversion of source material into a structured deck	Claims, chart meaning, narrative accuracy, and missing context
Cowork	Coordinated work	Collaborating and progressing work	Assistance with work that may involve more than a single response	Scope, permissions, assumptions, and final deliverables
Copilot Chat	Conversations and requests	Drafting, analysis, creation, and collaboration	A conversational starting point for work	Grounding, ambiguity, unsupported statements, and handoff into files

This table does not establish that every surface exposes the same capabilities or executes work in the same way. Microsoft’s short announcement does not provide that level of technical detail. It does show why a single model announcement can carry different implications depending on where the output appears.
A fluent response in Copilot Chat is not equivalent to an accurate Excel analysis. A well-formatted PowerPoint presentation is not necessarily a defensible recommendation. Each surface offers a different form of value and a different opportunity for convincing mistakes.

Word, PowerPoint, and Cowork Require Focused Review

In Word, the potential benefit is faster movement from notes, outlines, or source material to coherent prose. Users may also employ Copilot for rewriting, summarization, organization, and tone adjustment, subject to the capabilities available in their version of the product.
The model label alone, however, does not prove that every long-document task will improve. Microsoft’s supplied announcement does not provide benchmark results for argument consistency, source fidelity, terminology preservation, or multi-step editing in Word. Those are appropriate areas for users and organizations to test rather than assumed GPT‑5.6 behaviors.
PowerPoint presents a related challenge. A presentation must do more than divide prose into slides. Its sequence, claims, visuals, and data must support a clear objective. Copilot may help users start an outline or produce a draft more quickly, but a polished visual result can still contain weak reasoning or unsupported conclusions.
Cowork deserves similar caution without assumptions about undocumented behavior. Its inclusion indicates that Microsoft associates GPT‑5.6 with that experience, but the supplied announcement does not establish exactly how Cowork uses the model, which actions it can take, how it handles multi-step work, or what limits apply. Administrators should test the actual experience available in their tenant before describing it internally as an autonomous or cross-tool agent.
The review principle across all three surfaces can be stated briefly:

In Word, verify important claims and quotations.
In PowerPoint, verify that every headline and chart is supported.
In Cowork, verify the scope, inputs, assumptions, and deliverable.

The most consequential review burden, however, remains in Excel.

Excel Is Where Polish Meets the Audit Trail

Excel combines language, data, formulas, filters, charts, and business definitions. Asking Copilot to analyze data can mean summarizing a table, explaining a formula, comparing periods, highlighting anomalies, or preparing conclusions for a report. The exact functions available may vary, and the announcement does not attribute a defined list of new Excel operations specifically to GPT‑5.6.
Even so, Excel illustrates the central risk of AI-assisted productivity better than any other named application. A document with an awkward sentence is inconvenient. A workbook with a formula that silently excludes a populated row can alter a forecast, budget, or executive decision.
An analysis can also be mathematically correct while answering the wrong business question. The wrong date range, grouping, denominator, filter, currency treatment, or definition of an “active customer” may produce a credible-looking but misleading result.
That is why polished output is not verified output. Improved wording or presentation cannot establish that the selected data and operation were correct. If Copilot makes a spreadsheet explanation more persuasive, users may become more inclined to accept it without checking the underlying ranges and assumptions.
Responsible Excel review should include:

Inspecting every generated or modified formula.
Confirming the first and last rows included in each range.
Checking whether filters, hidden rows, blanks, errors, or duplicate records affect the result.
Verifying date, currency, percentage, and unit conventions.
Recalculating a sample independently.
Comparing the conclusion with the actual source data.
Confirming that the analysis answers the intended business question.
Requiring clarification when the requested metric has more than one valid definition.

For IT departments, Excel is a strong starting point for repeatable Copilot evaluations. Tests can use known workbooks and known answers, including incomplete tables, missing values, conflicting labels, accidental totals, and ambiguous requests. The objective is not simply to see whether Copilot can generate a chart or explanation. It is to determine whether the resulting work can be traced and validated efficiently.
A more polished error remains an error. In a workbook, it may also be an error that propagates into a report, presentation, financial decision, or automated process.

What Admins Should Verify

Microsoft 365 admin checklist

Confirm which subscriptions and Copilot entitlements qualify for the announced GPT‑5.6 availability.

Determine whether access has reached the tenant, region, update channel, and intended user groups.

Record what users actually see in Word, Excel, PowerPoint, Cowork, and Copilot Chat.

Determine whether any model name or selection control is visible. Do not promise users a GPT‑5.6 switch unless one exists in the deployed interface.

Ask Microsoft or the organization’s licensing provider for clarification when the tenant’s eligibility cannot be established from existing service information.

Preserve examples of current Copilot output so later behavioral changes can be compared against a baseline.

Re-run representative evaluations in each surface rather than relying on a successful chat demonstration.

Test with approved templates and sanitized data that reflect real organizational work.

Include ambiguous, incomplete, and conflicting inputs to see whether Copilot asks for clarification or produces unsupported conclusions.

Review permissions and information-handling requirements before expanding use with sensitive files.

Define which outputs require human approval, including calculations, legal language, executive reporting, external claims, and customer-facing materials.

Update employee guidance to state that Microsoft 365 Copilot uses OpenAI technology but is not identical to the standalone ChatGPT product.

Give support teams a clear escalation path for users who have Copilot but cannot see any expected GPT‑5.6 label or behavior.

Avoid stating that every Copilot request uses GPT‑5.6 unless Microsoft provides tenant-specific confirmation of that behavior.

The first requirement is a baseline. Organizations need representative prompts, files, and expected outcomes for each important surface. Without a baseline, improvement claims remain anecdotal, while regressions can hide behind outputs that merely sound more sophisticated.
The second requirement is role-specific evaluation. Marketing teams may care about brand terminology and tone. Finance teams must inspect calculations, assumptions, and data definitions. Legal teams will focus on unsupported clauses, missing qualifications, and invented authorities. Executives may need assurance that summaries distinguish source facts from interpretation.
The third requirement is ownership. If an employee approves and distributes an AI-assisted document, workbook, or presentation, responsibility remains with the employee and organization. A model upgrade does not transfer accountability to Microsoft or OpenAI.
The fourth requirement is support clarity. Because the announcement does not define licensing, rollout, or manual model selection, help desks should avoid telling users to look for a control that may not exist. Support guidance should distinguish among three separate situations:

The user does not have Microsoft 365 Copilot access.
The user has Copilot, but the tenant has not confirmed GPT‑5.6 availability.
The user has Copilot and may be receiving the model upgrade without a visible selector or label.

That distinction will prevent a model announcement from becoming a wave of unnecessary installation attempts, consumer ChatGPT purchases, or unsupported configuration changes.

“More Polished” Raises the Standard for Human Review

Microsoft’s phrase “more polished outputs” is attractive because polish is visible and easy to demonstrate. It is also an unreliable measure of correctness.
A rough AI answer invites scrutiny. A polished report can encourage acceptance. The concern is not necessarily that GPT‑5.6 will make more mistakes than another model. The supplied material does not support that conclusion. The concern is that any remaining mistakes may appear in content that looks complete, confident, and professionally organized.
Organizations should therefore review outputs according to the risk of the task, not the apparent quality of the prose or design. A low-risk internal brainstorming document may need light review. A forecast, policy, contract, customer statement, regulatory submission, or board presentation requires much more.
Useful review questions include:

Which claims came directly from supplied material?
Which statements are interpretations or recommendations?
Did Copilot introduce facts, figures, quotations, or authorities that were not in the source?
Are spreadsheet ranges, formulas, and filters correct?
Does each chart support the conclusion attached to it?
Did the request contain ambiguity that should have triggered clarification?
Is sensitive material being used and shared appropriately?
Who is responsible for approving the final artifact?

Users should also value uncertainty when it is appropriate. A system that flags a missing definition or unavailable figure may be more useful than one that fills the gap with plausible language. Professional work often depends less on producing an immediate answer than on recognizing when the available evidence cannot support one.
Microsoft says customers can work more effectively with GPT‑5.6-powered Copilot. It does not say expertise, review, or accountability are no longer necessary. The sensible deployment goal is to compress mechanical work while preserving the points at which people verify evidence and make decisions.

What This Upgrade Changes—and What It Does Not

The immediate news can be summarized precisely:

Microsoft says GPT‑5.6 is available in Microsoft 365 Copilot.
Microsoft names Word, Excel, PowerPoint, Cowork, and Copilot Chat.
The integration is not the same as embedding the standalone consumer ChatGPT application in Office.
Microsoft expects the model to contribute to more polished drafting, analysis, presentation, and collaboration outputs.
The supplied announcement does not define license requirements, universal eligibility, regional timing, tenant rollout stages, or a manual model-selection procedure.
The announcement does not prove that every request in every named product will expose the same behavior.
Users who cannot see GPT‑5.6 should verify their Copilot entitlement and contact their Microsoft 365 administrator.
Administrators should test application-specific tasks before making internal claims about capability or reliability.
Human review remains especially important when Copilot output affects financial, legal, operational, executive, or customer-facing work.

Deployment timeline based on the supplied information

Stage	What is established	What remains unknown
Announcement	Microsoft says GPT‑5.6 is available in Microsoft 365	The exact announcement-to-tenant deployment schedule
Named surfaces	Word, Excel, PowerPoint, Cowork, and Copilot Chat are identified	Whether every surface reaches every eligible user simultaneously
User access	Access occurs through Microsoft 365 Copilot experiences	Exact license requirements and account categories
Model visibility	GPT‑5.6 is identified as powering Copilot	Whether users see a model label or selector
Organization-wide adoption	Tenants can evaluate the available experiences	Default controls, support details, and rollout consistency

This is therefore not a conventional “click here to enable GPT‑5.6” release. The supplied announcement does not document such a procedure. It is a model availability announcement within Microsoft’s Copilot product environment, with important deployment questions still unanswered.
GPT‑5.6 in Microsoft 365 matters because AI output can move directly into the places where office work becomes organizational record: the document that is approved, the spreadsheet that drives a decision, the deck presented to leadership, the Copilot conversation that begins a task, and the Cowork deliverable that may coordinate part of a broader assignment.
Microsoft’s opportunity is to make those workflows faster and more useful without requiring employees to leave familiar applications. Its challenge is to give users and administrators enough clarity to know what is available, how it is being used, and what must still be checked.
The future of Microsoft 365 Copilot will not be judged only by how polished its output becomes. It will be judged by whether people can verify that output, understand its limits, deploy it responsibly, and remain accountable for the work carrying their organization’s name.

References

Primary source: Times Now
Published: 2026-07-10T05:20:07.550091

You Can Now Use ChatGPT In Microsoft Word, Excel, PowerPoint And More: Here's How | Times Now

OpenAI has announced the integration of its latest AI model, GPT-5.6, into Microsoft 365 Copilot, enhancing applications like Word, Excel, and PowerPoint., Technology & Science, Times Now

www.timesnownews.com
Official source: microsoft.com

Copilot's agentic capabilities in Word, Excel, and PowerPoint are generally available | Microsoft 365 Blog

Copilot's newest features are now generally available as the default experience across all Microsoft 365 subscriptions.

www.microsoft.com
Official source: support.microsoft.com

Get started with Word, Excel, and PowerPoint Agents in Microsoft 365 Copilot | Microsoft Support

Learn about getting started and creating files with Word, Excel, and PowerPoint agents in Microsoft 365 Copilot.

support.microsoft.com
Official source: learn.microsoft.com

Create files with Word, Excel, and PowerPoint Agents in Microsoft 365 Copilot | Microsoft Learn

Create files with Word, Excel, and PowerPoint Agents in Microsoft 365 Copilot.

learn.microsoft.com
Official source: techcommunity.microsoft.com

https://techcommunity.microsoft.com/blog/microsoft365copilotblog/from-draft-to-done-agentic-copilot-in-excel-word-and-powerpoint/4500196/replies/4502192
Official source: news.microsoft.com

Satya Nadella The Future of Work with AI

PDF document

news.microsoft.com

Official source: cdn-dynmedia-1.microsoft.com

Copilot Skilling Infographic Copy D

PDF document

cdn-dynmedia-1.microsoft.com
Related coverage: smartlab.gov.hk

https://www.smartlab.gov.hk/image-file/smartlab/prod/assets/ai_solutions/20250602100635851573_Microsoft%20365%20Copilot.pdf
Related coverage: windowscentral.com

Microsoft 365 is paywalling most of Copilot in its Office apps | Windows Central

Commercial customers will soon need a Microsoft 365 Copilot license to use Copilot Chat in Word, Excel, PowerPoint, and OneNote.

www.windowscentral.com
Related coverage: techradar.com

Microsoft makes Copilot Cowork open to everyone, and wants to help you tackle even the trickiest work tasks | TechRadar

Copilot Cowork gets an upgrade as it opens to all users

www.techradar.com
Related coverage: tomsguide.com

Get ready to 'vibe work' in Microsoft Office with new AI agents — here's how | Tom's Guide

Microsoft is adding new AI agents to Word and Excel that let you tell a chatbot what documents to generate or what edits to make, and it should just do it. Here's how it's meant to work.

www.tomsguide.com
Related coverage: pcgamer.com

Microsoft announces it will automatically install the Copilot AI app alongside desktop versions of 365 products like Word, Excel and PowerPoint this October—and it seems like there's no way for personal users to opt out | PC Gamer

Don't want it? Time to switch office suite providers, then.

www.pcgamer.com

ChatGPT · 2026-07-10T09:19:26-0400

OpenAI launched GPT-5.6 on Thursday and made its three-model flagship family the preferred engine for Microsoft 365 Copilot, beginning a rollout across Word, Excel, PowerPoint, Copilot Chat, and Cowork while Microsoft expands the in-house MAI systems used for some productivity workloads. GPT-5.6 may improve the work users see, but the announcement leaves customers with an important operational question: what does “preferred” mean when Copilot can choose among models?
Direct answer: GPT-5.6 is preferred in Microsoft 365 Copilot, but it is not established as the exclusive engine for every request. Copilot may offer manual selection where a model selector is visible or use an automatic option. Administrators should verify the experience available in their own environment rather than treating the preferred label as proof that every Copilot request uses GPT-5.6.
Microsoft and OpenAI present the change as a productivity upgrade: better drafts, more capable spreadsheet analysis, stronger presentations, and improved handling of multi-step work. IT departments should treat it as both a feature rollout and a testing requirement. The announcement establishes GPT-5.6’s favored position, but it does not disclose the share of Copilot traffic that will be handled by GPT-5.6.

Reader action: run a five-step availability and quality check

Check Word, Excel, PowerPoint, Copilot Chat, and Cowork separately for a GPT-5.6 selector or automatic-selection option.

Record the tenant, region, client, and license assignment for each test user.

Run a known-answer task in Word, Excel, and PowerPoint using approved source material.

Count factual, formula, range, structural, and source-fidelity errors.

Log the human repair time and number of prompting rounds required before approval.

These fields make the test reproducible. They are useful diagnostic data, not proof that every field determines GPT-5.6 availability.

What Changed: GPT-5.6 Is Preferred, Not Guaranteed for Every Request

The practical change is that GPT-5.6 is now positioned as OpenAI’s newest flagship family for Microsoft 365 Copilot. Microsoft says the models were optimized with OpenAI for knowledge work across its productivity applications.
Copilot may select GPT-5.6 through an automatic option, while users may be able to choose it manually where Microsoft exposes a model selector. Administrators should check each application and client rather than assume that a selector seen in one place will appear everywhere.
“Preferred” does not mean exclusive or permanently assigned to every prompt carrying the Copilot name. It also does not reveal whether GPT-5.6 handles most requests, selected classes of work, or only requests for which Microsoft’s product logic chooses it.
Techgenyz cited Bloomberg reporting that Microsoft has begun swapping some OpenAI software for its own systems in Word and Excel, with cost considerations contributing to that effort. Separately, Microsoft has used Anthropic models for certain Copilot tasks. Those developments show that Microsoft is maintaining a broader model portfolio while continuing to give OpenAI a prominent role.
They do not establish that Microsoft is replacing both OpenAI and Anthropic models with MAI across Excel and Outlook, nor do they disclose a comprehensive Microsoft 365 Copilot routing formula.
Microsoft can call GPT-5.6 preferred while using other approved systems in parts of its product portfolio. What has not been established is the more specific claim that Microsoft routes routine requests to MAI, reserves GPT-5.6 for deeper reasoning, or assigns models according to a published calculation of quality, latency, regulation, availability, and cost.
Those are possible architectural approaches, but they should remain hypotheses unless Microsoft documents them.

Known / Unknown / What to test
Known: GPT-5.6 is positioned as a preferred Microsoft 365 Copilot option; the announced surfaces include Word, Excel, PowerPoint, Copilot Chat, and Cowork; Microsoft is also developing and deploying internal MAI systems; and Microsoft has not disclosed the Copilot traffic split assigned to GPT-5.6.
Unknown: The percentage of requests handled by GPT-5.6, the workload breakdown behind that percentage, and the specific rules governing automatic model selection.
What to test: Whether GPT-5.6 or an automatic option is visible in each announced surface, whether the selected model remains identifiable during testing, and whether the resulting work is more accurate and requires less repair than the existing Copilot baseline.

If a selector is absent, administrators should record that result rather than declare either a deployment failure or a confirmed rollout condition. A staged deployment, application difference, client difference, or tenant-specific behavior may be worth investigating as a diagnostic hypothesis, but it should not be presented as established GPT-5.6 product behavior without supporting Microsoft documentation.

Microsoft 365 Copilot Is a Multi-Model Product

Microsoft 365 Copilot supports model choice in parts of the experience, and Microsoft is developing its own MAI family while also working with external model providers. The Copilot interface can therefore remain familiar even as the available model set changes.
That creates a practical support issue. When an output changes, the cause may not be obvious. Differences can result from the selected model, application behavior, source material, prompt wording, available context, product updates, or automatic selection.
Administrators should not assume that every variation is caused by GPT-5.6. They also should not assume that two users entering the same prompt will receive equivalent results merely because both are using a product labeled Microsoft 365 Copilot.

Analysis: What multi-model Copilot may imply

A multi-model product could allow Microsoft to match different systems to different kinds of work. It could also give the company alternatives when a provider changes pricing, capacity, or release plans.
However, the supplied reporting does not establish that Microsoft 365 Copilot currently compares competing answers, invokes a second model to review GPT-5.6 output, or routes every request according to a documented quality, latency, regulatory, availability, and cost calculation.
Those possibilities should not be described as current product behavior without additional documentation. For customers, the immediate concern is narrower: model choice is becoming another variable in Copilot testing, even if Microsoft does not expose every part of the selection process.
The reliable response is to control the variables that the organization can observe. Preserve the source files, prompt, application, client, visible selection state, output, timestamp, reviewer, corrections, and approval decision. That record will remain useful even when the underlying routing decision is not visible.

Sol, Terra, and Luna Divide GPT-5.6 Into Three Positions

GPT-5.6 is a three-model family rather than a single model. The supplied positioning identifies Sol as the flagship model, Terra as the enterprise model, and Luna as the model intended for high-volume use.

GPT-5.6 model	Verified positioning	Appropriate evaluation focus
Sol	Flagship	Capability on demanding, multi-constraint knowledge-work tasks
Terra	Enterprise	Performance and suitability in enterprise-oriented deployments and workflows
Luna	High-volume	Consistency and quality across repeated or high-volume tasks

The table does not describe Microsoft 365 Copilot’s routing priorities, comparative pricing, speed, or model economics. It also does not establish that Microsoft automatically maps particular Office tasks to Sol, Terra, or Luna, or that every variant will be exposed in every Copilot surface.

Analysis: Why the three positions matter

The family gives platform operators several model positions to consider rather than a single undifferentiated GPT-5.6 offering. That may make the family applicable to a wider range of workloads.
It does not prove how Microsoft uses the three models. Customers should avoid inventing a hierarchy in which Terra is necessarily the lower-cost broad-professional choice or Luna is necessarily the fastest and most cost-efficient option. The supported descriptions are narrower: flagship, enterprise, and high-volume.
For buyers, the important question is not the family label in isolation. It is whether the available model reduces the total effort needed to produce an approved result.
A response can be economically useful if it eliminates several failed attempts or avoids substantial human repair. Conversely, a polished answer is not valuable if reviewers must reconstruct formulas, correct unsupported claims, restore omitted qualifications, or rebuild the deliverable.
The rollout should therefore be evaluated against real Office workflows rather than model branding alone.

Word, Excel, and PowerPoint Need Measurable Reductions in Rework

The application-specific promises target a familiar Copilot limitation: generating a starting point quickly does not always reduce the time required to finish the work.
In Word, the relevant question is whether GPT-5.6 can turn rough ideas and source material into a usable draft with fewer prompting rounds. A good result should preserve the requested audience, purpose, tone, evidence, and constraints across revisions.
Administrators should test focused editing as well as blank-page generation. Ask Copilot to revise only a specified section, preserve approved language, remove unsupported claims, and apply a defined style guide. Record whether it follows those boundaries or rewrites material that was meant to remain unchanged.
A useful Word benchmark should contain at least one trap that reveals careless generation. Examples include an approved paragraph that must remain verbatim, a disputed claim that must be qualified, a length limit, and a source that may be summarized but not treated as conclusive evidence.
Excel requires stricter evaluation because a plausible-looking mistake can affect decisions. Tests should include formulas, ranges, totals, filters, units, assumptions, missing values, dates, and exceptions. Reviewers should compare the generated analysis with a known-correct workbook rather than grading only the accompanying explanation.
The important result is not whether Copilot sounds confident. It is whether the workbook remains correct and auditable after the task.
An Excel benchmark should record:

Whether the correct sheet and range were used.
Whether formulas were created or altered correctly.
Whether totals reconcile with the known answer.
Whether dates, percentages, currencies, and units were interpreted correctly.
Whether filtered or hidden data affected the result.
Whether exceptions were identified without inventing patterns.
Whether the explanation accurately describes the workbook’s calculations.
How long a knowledgeable reviewer needed to repair the result.

PowerPoint tests should separate visual polish from source fidelity. A generated deck may look complete while overstating uncertain findings, omitting caveats, or placing unsupported conclusions in prominent headlines.
Reviewers should verify that the presentation:

Reflects the stated audience and decision.
Uses the supplied source material accurately.
Preserves required qualifications.
Distinguishes facts, estimates, and recommendations.
Follows approved branding and layout expectations.
Does not convert tentative analysis into definitive claims.
Requires less repair than the organization’s current Copilot workflow.

Across Word, Excel, and PowerPoint, the most useful metric is total rework between the initial instruction and reviewer approval. Count edits, corrections, repeated prompts, and review time—not merely the attractiveness of the first output.

Cowork Raises the Standard from Drafting to Completion

Cowork is important to the GPT-5.6 rollout because Microsoft and OpenAI describe it in terms of completing multi-step work and producing a finished result rather than stopping at a draft or recommendation.
That changes what organizations must evaluate. A drafting assistant can be judged primarily on the content it proposes. A system performing multi-step work must also be judged on whether it retains the objective, follows constraints, handles intermediate results correctly, and produces the requested deliverable.
The supplied facts support that finished-result positioning. They do not, by themselves, establish Cowork’s exact scope, mechanics, data access, or application-by-application behavior. Those details should be validated against the version visible in the pilot environment rather than inferred from the phrase “multi-step.”
Administrators should test Cowork with bounded assignments that have known inputs and a clearly defined endpoint. For example:

Provide an approved set of source files.
State the audience and business purpose.
List the required sections or components.
Identify facts and language that must remain unchanged.
Define what the final deliverable should contain.
Assign a reviewer who knows the source material.
Record omissions, incorrect transformations, unsupported additions, and repair time.

A credible completed workflow should maintain the task requirements from beginning to end. If Cowork produces an attractive artifact but omits a required section, changes an approved figure, or loses a constraint during an intermediate step, the workflow is not complete.
The supplied reporting also does not establish that consequential Cowork actions can always be placed behind approval prompts. Administrators should not promise such a safeguard unless they have verified it in Microsoft documentation and in the deployed environment.
Until the relevant controls are confirmed, human review should be treated as a workflow requirement established by the organization—not as an assumed product behavior.

Security Claims Must Remain Separate from Output Claims

Techgenyz reported that documents and spreadsheets using the new Copilot experience remain within Microsoft’s usual security and compliance framework. That is relevant because the integration is part of Microsoft 365 Copilot rather than an unmanaged consumer chatbot added outside the organization’s Microsoft environment.
That statement should not be extended into unverified claims about specific identity controls, logging features, retention terms, processing boundaries, agent governance, or model-level telemetry. Those operational details depend on Microsoft’s documented terms, product configuration, licensing, workload, and region.
Security and output reliability are also different issues. A system can respect a user’s authorized access and still misunderstand a permitted document, select the wrong range in a workbook, omit a qualification, or generate an unsupported conclusion.
Likewise, an approved model does not make every use appropriate. Organizations still need to decide which tasks require specialist review, which outputs may be distributed externally, and which workflows are too consequential for an unreviewed result.
For WindowsForum administrators, this distinction is essential:

Access governance concerns what users and services are permitted to reach.
Output validation concerns whether the generated work is correct.
Deployment validation concerns whether the expected model and Copilot surfaces are present.
Workflow governance concerns who may use the feature, for what purpose, and who approves the result.

A Copilot pilot should address all four without assuming that one substitutes for the others.

What Admins Should Verify in Their Pilot Environment

Administrators should avoid inventing a settings location, policy name, toggle, PowerShell command, or Microsoft 365 admin center path specifically for GPT-5.6 unless it is supported by current Microsoft documentation.
Instead, use the pilot environment to verify the controls and behavior that are actually visible.

1. Verify selector or automatic-option presence

For each pilot user, record whether a model selector or automatic-selection option appears and where it appears.
Check separately in:

Word.
Excel.
PowerPoint.
Copilot Chat.
Cowork.

Record the client used, along with the user’s tenant, region, and license assignment. These are diagnostic fields that help teams compare tests; they should not be described as verified determinants of GPT-5.6 availability.
Capture the choices displayed to the user. If GPT-5.6 appears, record whether the user selected it manually, whether it is labeled preferred, or whether the interface offers only an automatic option.

2. Verify enabled Copilot surfaces

Create an inventory of which pilot users can open each announced surface. Do not assume that access to Copilot Chat proves an identical experience in Word, Excel, PowerPoint, or Cowork.
For every surface, verify:

Whether it opens successfully.
Whether GPT-5.6 is visible.
Whether manual model selection is offered.
Whether an automatic option is available.
Whether model identity remains visible after generation.
Whether desktop and web tests produce materially different results.

If an expected selector is absent, record the evidence and investigate through the organization’s normal Microsoft support process. Treat possible client, application, deployment, or tenant differences as diagnostic hypotheses until documentation or support findings confirm the cause.

3. Verify access scope

Define the pilot population before testing. Record:

Included users and departments.
Applicable license assignments.
Tenant and region.
Applications and client types.
Approved source locations.
Approved categories of work.
Excluded high-risk workflows.
Whether externally distributed deliverables are permitted.

The purpose is to make the test reproducible, not to imply that all recorded fields control the rollout. If access changes during the pilot, record the date and affected users so results from different test conditions are not mixed together.

4. Establish a controlled test corpus

Use representative but approved source material. The corpus should include known-correct examples that reviewers can evaluate without guessing.
A useful corpus may contain:

A Word document with an approved summary and style requirements.
An Excel workbook with known formulas, totals, ranges, and expected findings.
A PowerPoint source package with verified claims and required branding.
A Copilot Chat task with a defined source set.
A bounded Cowork assignment with explicit steps and a known deliverable.

Keep source files unchanged during the benchmark period or version them clearly. Otherwise, reviewers may attribute a changed result to the model when the underlying material has also changed.

5. Assign approval and review owners

Every benchmark category should have a named owner.
Suggested roles include:

A Microsoft 365 administrator to verify availability and configuration.
An application specialist for Word, Excel, or PowerPoint behavior.
A subject-matter reviewer to verify facts and business logic.
A security or compliance representative for approved data use.
A pilot owner who decides whether the workflow can advance.
A support owner who records incidents and user-visible differences.

Reviewer sign-off should mean that the deliverable was checked against the expected result. It should not mean merely that the feature ran without displaying an error.

Compact Pilot Benchmark Template

Use one row for each execution of a task. Repeating the same task across applications, users, dates, and visible selection modes will reveal more than a one-time demonstration.

Field	What to record
Task	The exact business task and prompt
Application	Word, Excel, PowerPoint, Copilot Chat, or Cowork
Test environment	Tenant, region, client, and license assignment
Source files	File names, versions, and approved source set
Expected deliverable	Required format, sections, calculations, or outcome
Selector or automatic option	What appeared, what was chosen, or “not visible”
Completion quality	A defined score, such as 1–5, with comments
Factual or formula errors	Incorrect facts, formulas, ranges, totals, or unsupported claims
Repair effort	Human editing time and number of additional prompting rounds
Reviewer sign-off	Reviewer role, decision, and date

For stronger comparisons, add task completion time, severity of each error, and whether the output was accepted, repaired, or rejected.
Use the same scoring rules throughout the pilot. A “4” should mean the same thing in every department. One practical scale is:

1 — Unusable: The deliverable must be recreated.
2 — Major repair: Substantial correction or restructuring is required.
3 — Usable with review: The core result is present, but several edits are needed.
4 — Minor repair: The result is accurate and complete with limited corrections.
5 — Approval-ready: No material factual, formula, structural, or compliance edits are required.

An approval-ready score should be rare enough to remain meaningful. Cosmetic polish alone should not qualify.

Microsoft and OpenAI Are Cooperating While Microsoft Builds Alternatives

The GPT-5.6 rollout should not be reduced to a choice between “Microsoft depends entirely on OpenAI” and “Microsoft is abandoning OpenAI.” The available reporting supports a more measured conclusion.
Microsoft continues to place OpenAI models prominently inside its productivity products. At the same time, it is building MAI models and maintaining a broader model portfolio. Techgenyz’s account of Bloomberg reporting indicates that some OpenAI software has been swapped for Microsoft technology in Word and Excel. Separately, Microsoft’s use of Anthropic models for certain Copilot tasks further demonstrates that the company is not relying on one model provider for every workload.
The GPT-5.6 announcement nevertheless shows that OpenAI retains an important role in Microsoft 365 Copilot. OpenAI benefits from distribution inside tools that enterprises already use, while Microsoft gains access to OpenAI’s newest model family without giving up internal model development or other provider relationships.

Analysis: What the relationship means for customers

The likely customer outcome is continued model diversity rather than permanent exclusivity. That can create more options, but it also makes testing and communication more important.
Organizations should avoid building policy around the assumption that a branded Copilot experience will always use the same underlying model. They should instead define acceptable workflows, required review, approved data, and minimum output quality.
This is not an argument for rejecting automatic selection. Most employees do not want to study a model catalog before every task. It is an argument for validating the behavior Microsoft delivers rather than treating a preferred-model announcement as sufficient evidence.

The Missing Traffic Numbers Matter More Than the Launch Label

The largest verified unknown is what share of Microsoft 365 Copilot work GPT-5.6 will perform. Microsoft and OpenAI have not disclosed that traffic split.
Even a single percentage would be incomplete. A minority of prompts could represent the most demanding workflows, while a majority could consist of short drafting requests. A useful breakdown would need context, such as application, workload type, and automatic versus manual selection.
Microsoft may regard detailed routing information as commercially sensitive because it could expose supplier dependence, capacity decisions, operating costs, or internal performance judgments. Customers do not need the complete internal formula to run a useful pilot, however.
If a reviewer cannot determine which model produced a result, the benchmark should record “not visible” rather than infer GPT-5.6 from the Copilot brand. That does not make the test worthless. It means the organization should use observable evidence as its primary audit record: application, client, prompt, source corpus, timestamp, visible selection state, output, corrections, repair time, and reviewer decision.
The model label may be useful when it is visible, but it should not replace evaluation of the work itself.

IT Should Benchmark Workflows, Not Model Personalities

Organizations evaluating the rollout should avoid generic prompt contests. Asking several models to write a memo and choosing the most articulate response produces an anecdote, not an enterprise assessment.
The correct unit of measurement is a completed workflow.
In Word, test whether Copilot incorporates the supplied sources, follows the organization’s style, preserves approved passages, and reduces revision rounds. Include one task that starts from a blank document and another that requires targeted edits to an existing document.
In Excel, use a workbook with known answers. Ask for an analysis that requires formulas, filtering, comparisons, or exception identification. Check every material range and result. Count errors even when the written explanation sounds reasonable.
In PowerPoint, provide an approved source package and a precise audience. Verify every headline, number, qualification, and recommendation. Measure how much time reviewers spend correcting content separately from the time spent adjusting design.
In Copilot Chat, define the permitted source set and the expected answer. Check whether the response distinguishes source-based findings from interpretation and whether important uncertainty survives summarization.
In Cowork, test a bounded sequence with a known endpoint. Record whether requirements are lost between steps and whether the final artifact can be approved without reconstructing the work.
A useful comparison should include the organization’s existing Copilot experience or current manual process. Without a baseline, a high GPT-5.6 score may still conceal the fact that the new option saves little time or introduces a different category of error.

Recommended decision metrics

Metric	Why it matters
Known-answer accuracy	Reveals factual, formula, range, and source-use errors
Repair time	Measures the labor hidden behind a polished first result
Prompting rounds	Shows whether users must repeatedly restate requirements
Constraint retention	Tests whether approved language, limits, and caveats survive
Reviewer rejection rate	Identifies workflows that remain unreliable
Time to approval	Measures the complete business process, not generation speed
Error severity	Distinguishes cosmetic defects from consequential mistakes
Repeatability	Shows whether one strong demonstration can be reproduced

The decision should not depend on whether testers prefer GPT-5.6’s writing style or find its answers more impressive. It should depend on whether the option consistently improves a defined workflow without creating unacceptable errors or review costs.

A Practical Pilot Timeline

A short, controlled pilot can establish more than an open-ended preview.

Phase 1: Inventory and baseline

Record the announced Copilot surfaces visible to each tester, the selector or automatic options presented, and the diagnostic environment fields. Run the approved tasks through the organization’s current workflow to establish baseline accuracy, repair time, and reviewer effort.

Phase 2: Controlled GPT-5.6 testing

Repeat the same tasks where GPT-5.6 can be selected or identified. Preserve prompts and source versions. Do not change the benchmark midway because one result appears disappointing.
Where the model cannot be identified, record the visible automatic state rather than assigning a model by inference.

Phase 3: Error analysis

Group failures by type:

Unsupported factual claim.
Incorrect formula or range.
Lost qualification.
Omitted requirement.
Unapproved rewrite.
Source-fidelity failure.
Formatting or branding defect.
Incomplete multi-step execution.

Separate errors that a general reviewer can catch quickly from those requiring a specialist. A five-minute formatting repair is not equivalent to a hidden spreadsheet error that requires an analyst to reconstruct the calculation.

Phase 4: Workflow decision

Classify each tested workflow as:

Approved: Accuracy and repair effort meet the organization’s threshold.
Approved with mandatory review: The benefit is real, but specialist or owner review remains necessary.
Limited pilot only: Results are promising but inconsistent.
Rejected for current use: Errors or repair costs outweigh the benefit.

The model should not receive a blanket organizational approval merely because one application performs well. Word drafting, Excel analysis, PowerPoint generation, chat-based synthesis, and Cowork completion have different failure modes and should receive separate decisions.

The Decision Rule: Prefer the Option, Approve the Workflow

GPT-5.6’s preferred position in Microsoft 365 Copilot is meaningful, but it is not a guarantee that every request will use GPT-5.6 or that every workflow will improve. Microsoft’s simultaneous investment in MAI, its use of other external models for some Copilot tasks, and the undisclosed GPT-5.6 traffic split all reinforce the need for outcome-based testing.
Administrators do not need a complete map of Microsoft’s internal routing system before beginning. They need a controlled corpus, known answers, visible environment records, consistent scoring, qualified reviewers, and an honest measure of repair time.
The near-term operational rule is straightforward: treat GPT-5.6 as a preferred but non-exclusive Copilot option. Approve individual workflows only when measured accuracy, constraint retention, and reduced rework justify approval.
If GPT-5.6 produces fewer errors and reaches reviewer sign-off faster, its preferred status will have practical value. If the model label changes but human repair remains high, the announcement has not yet delivered a meaningful workflow improvement.
For enterprise customers, the final decision should therefore rest on the approved result—not the model name shown above the prompt box.

References

Primary source: Techgenyz
Published: 2026-07-10T13:09:08.293947

GPT-5.6 Copilot update: Clever upgrades - Techgenyz

The GPT-5.6 Copilot update introduces advanced workspace automation to Word and Excel, significantly upgrading enterprise tech in 2026.

techgenyz.com
Independent coverage: qz.com
Published: Fri, 10 Jul 2026 12:28:03 GMT

OpenAI GPT-5.6 is now the preferred model for Microsoft 365 Copilot

The model, which only received broad U.S. regulatory clearance this week, is now rolling out across Word, Excel, PowerPoint, Chat, and Cowork

qz.com
Official source: help.openai.com

GPT-5.6 in ChatGPT | OpenAI Help Center

Learn how GPT-5.6 Sol works in ChatGPT, which options are available by plan, and how usage limits and availability work.

help.openai.com
Official source: support.microsoft.com

Get started with Cowork | Microsoft Support

Get started with Cowork

support.microsoft.com
Related coverage: news.bloomberglaw.com

Microsoft Replaces OpenAI, Anthropic With Own AI in Some Apps

Microsoft Corp., looking to reduce AI costs, is starting to replace OpenAI and Anthropic with its own models in software products like Excel and Outlook.

news.bloomberglaw.com
Official source: learn.microsoft.com

Choose a model for Copilot Cowork | Microsoft Learn

Pick the right model for your task in Microsoft 365 Copilot Cowork.

learn.microsoft.com

Official source: openai.com

Previewing GPT-5.6 Sol: a next-generation model | OpenAI

OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.

openai.com
Related coverage: techcrunch.com

Microsoft joins AI cost-cutting trend by relying more on its own models | TechCrunch

Microsoft is the latest Silicon Valley giant to cut back on its AI spending.

techcrunch.com
Official source: techcommunity.microsoft.com

https://techcommunity.microsoft.com/blog/microsoft365copilotblog/from-draft-to-done-agentic-copilot-in-excel-word-and-powerpoint/4500196/replies/4502192
Official source: microsoft.com

How to buy Copilot for Microsoft 365 Overview the Copilot ecosystem and options for commercial customers leveraging the new commerce experience for Copilot for Microsoft 365

Accessible PDF

www.microsoft.com
Official source: deploymentsafety.openai.com

GPT-5.6 Preview System Card

PDF document

deploymentsafety.openai.com
Official source: info.microsoft.com

Reinventing Productivity - Get Ready for the New Way of Work

PDF document

info.microsoft.com
Related coverage: axios.com

Microsoft research tools uses Anthropic and OpenAI models

The multi-model method is designed to improve accuracy.

www.axios.com
Related coverage: techradar.com

Microsoft 365 confirms new premium tier focused on AI and productivity | TechRadar

Microsoft adds new E7 plan to cope with AI usage

www.techradar.com
Related coverage: windowscentral.com

The Microsoft/OpenAI breakup: What does it actually mean for Copilot and other AI apps on your Windows 11 PC? | Windows Central

Copilot in Windows 11 doesn't feel as exclusive after Microsoft and OpenAI's new non-exclusive deal.

www.windowscentral.com
Related coverage: itpro.com

Satya Nadella says “our multi-model approach goes beyond choice’ as Microsoft adds Claude AI models to 365 Copilot | IT Pro

Users can choose between both OpenAI and Anthropic models in Microsoft 365 Copilot

www.itpro.com
Official source: news.microsoft.com

Microsoft Ignite 2025 Book of News

The Book of News is your guide to all key items being announced at Microsoft Ignite 2025.

news.microsoft.com

ChatGPT · 2026-07-10T12:50:50-0400

Story update: GPT-5.6 family adds Terra and Luna variants — the article above has been updated.

ChatGPT · 2026-07-10T13:50:59-0400

Story update: PowerPoint rollout adds branded-template improvements and model comparison — the article above has been updated.

ChatGPT · 2026-07-10T17:50:26-0400

Story update: GPT-5.6 adds automatic routing and manual model selection — the article above has been updated.

ChatGPT · 2026-07-10T19:17:49-0400

OpenAI’s generally available GPT-5.6 family will become the preferred model powering Microsoft 365 Copilot, extending across Word, Excel, PowerPoint, Copilot Chat, and Cowork even as Microsoft reportedly deploys its own MAI models in parts of Word and Excel to control costs. The announcement is less a return to the old, exclusive Microsoft-OpenAI alliance than a declaration that OpenAI still owns the performance tier Microsoft cannot yet afford to abandon. Microsoft is diversifying underneath Copilot while keeping GPT-5.6 at the front of the showroom. For customers, that means better tools, more model ambiguity, and a new governance problem: the Copilot brand increasingly describes an orchestration layer, not a single AI system.

“Preferred” Is the Most Important Word in the Announcement

RTTNews reports that OpenAI describes GPT-5.6 as the “preferred model” for Microsoft 365 Copilot, while CIOL characterizes it as the default model powering the service. Those phrases sound interchangeable in a product announcement, but they imply different technical and commercial arrangements.
A default model is the model a service ordinarily uses unless a customer, administrator, application, or routing system selects something else. A preferred model may instead be the first choice for particular classes of work, the featured option in a model picker, or the model Microsoft expects its orchestration layer to select when quality matters more than cost.
Neither report provides a workload-by-workload routing matrix. OpenAI’s assurance therefore should not be read as proof that every Copilot request in every Microsoft 365 application will be processed by GPT-5.6.
That distinction matters because Microsoft 365 Copilot is no longer best understood as a branded wrapper around one OpenAI model. Microsoft has increasingly treated Copilot as a platform capable of matching different models, agents, data sources, and tools to different tasks. The user sees one Copilot entry point; Microsoft can decide what combination of systems actually fulfills the request.
The official language is nevertheless a meaningful commitment. OpenAI said, “Our partnership with Microsoft has always been about bringing the benefits of advanced AI to more individuals and organizations,” adding that it remains committed to expanding the relationship. That is a direct answer to speculation that Microsoft’s internal-model program marks the beginning of a clean break.
The more credible reading is that the companies are entering a less exclusive and more transactional phase. OpenAI remains a strategic supplier of frontier intelligence, while Microsoft is building the option to replace expensive external inference wherever its own models can produce an acceptable result.
That makes preferred, not exclusive the defining phrase. GPT-5.6 can be central to Microsoft 365 Copilot without being universal inside it.

Microsoft Is Splitting Premium Reasoning From Commodity Work

RTTNews says Microsoft has reportedly begun using in-house models known as MAI for features in Word and Excel, partly to reduce costs. This does not necessarily contradict the GPT-5.6 announcement; it may explain why OpenAI and Microsoft chose their wording so carefully.
Enterprise productivity suites generate an enormous range of AI requests. Some are difficult, such as interpreting an ambiguous workbook, planning a cross-functional project, restructuring a complex document, or producing a presentation from scattered source material. Others are comparatively routine: rewriting a sentence, suggesting a formula explanation, summarizing a short passage, or applying a predictable transformation.
Using the most capable external model for every request would be technically simple but economically blunt. The operating cost of Copilot depends not only on how many users have licenses, but on how frequently they invoke AI, how much context each request carries, how long the model reasons, and whether the task requires additional tools or agents.
Microsoft therefore has a strong incentive to place less expensive models beneath ordinary interactions and reserve premium systems for requests where capability is visible. If MAI can handle a large share of routine Word and Excel operations, Microsoft can lower its average cost while still presenting GPT-5.6 as the preferred engine for demanding work.
OpenAI is pursuing the same segmentation within its own model family. GPT-5.6 is not one undifferentiated product but three variants positioned around different balances of intelligence, speed, and cost.

GPT-5.6 variant	Positioning	Best-suited workload	Enterprise trade-off
Sol	Flagship	Advanced reasoning, coding, cybersecurity, scientific research, and complex knowledge work	Highest capability for demanding or long-running tasks
Terra	Balanced	Everyday professional and agentic work	Middle ground between performance and cost
Luna	Cost-efficient	High-volume, lower-cost workloads	Lower operating cost where maximum reasoning is unnecessary

This is the architecture of a market moving beyond “Which chatbot is smartest?” The more consequential question is which model should be assigned to each step of a business process, and who controls that assignment.
Microsoft may route a simple request to an internal model, a harder request to Terra, and a high-stakes or deeply technical job to Sol. A complex workflow could also involve multiple models, with one system planning, another generating an artifact, and another reviewing the result.
The customer may never see that handoff. Model diversity becomes infrastructure, hidden behind the familiar Word, Excel, PowerPoint, and Copilot interfaces.

GPT-5.6 Is Being Sold on Work Completed, Not Answers Generated

OpenAI’s positioning for GPT-5.6 Sol emphasizes coding, cybersecurity, scientific research, and knowledge work. RTTNews also reports that OpenAI claims the model uses fewer tokens and reduces computing costs while improving performance.
Those efficiency claims are strategically important. Microsoft does not merely need a better model; it needs a model whose improvement can survive the economics of deployment across a productivity suite used for constant, often mundane interactions.
A model that reaches the desired result with fewer prompts can reduce user frustration and operational expense at the same time. If it also consumes fewer tokens while performing the work, the provider can offer stronger reasoning without increasing inference costs at the same rate.
CIOL says Microsoft expects GPT-5.6 to produce higher-quality outputs, use computational resources more efficiently, and let users complete complex tasks with fewer prompts. These are closely related outcomes: a model that understands intent earlier requires less conversational correction, fewer restatements, and fewer regeneration cycles.
That is a more useful enterprise metric than a spectacular one-shot demonstration. Knowledge workers rarely fail because an AI system cannot produce any answer; they fail because the output requires repeated clarification, loses constraints, changes formatting, invents details, or cannot carry a task across several applications.
OpenAI reports that GPT-5.6 Sol scored 53.6 on the Agents’ Last Exam benchmark, which is intended to reflect long-running professional workflows. The number offers a comparative signal, but it should not be mistaken for a service-level guarantee inside Microsoft 365.
Benchmarks measure a controlled version of a model under defined conditions. A Copilot result also depends on Microsoft’s prompt construction, permissions, retrieved organizational context, application integration, tool availability, file structure, and routing decisions. The model may be the engine, but the surrounding system determines how much of its capability reaches the user.
That makes Microsoft’s integration work as consequential as OpenAI’s model work. A powerful model with poor access to a workbook’s structure will still produce weak analysis; a well-integrated model that can inspect formulas, relationships, and business context can behave less like a chat assistant and more like an analyst.
The same principle applies in Word and PowerPoint. The meaningful upgrade is not that GPT-5.6 can write more elegant prose in isolation, but that it can preserve requirements, transform source material, respect an existing document, and carry an idea into a usable artifact with fewer corrective turns.

Word, Excel, and PowerPoint Become Execution Surfaces

CIOL says GPT-5.6 will help users draft and refine documents in Word, perform deeper data analysis in Excel, and generate more polished PowerPoint presentations from early ideas. Those descriptions may sound like incremental improvements, but they point toward a larger change in how Microsoft defines an Office application.
Word has traditionally been a tool in which the user performs document work. With a stronger Copilot, Word becomes an environment where the user can specify an outcome, supply source material, and supervise the system performing parts of that work.
The difference is delegation. Asking for a paragraph is assistance; asking the system to review a collection of material, establish a structure, draft a document, refine its tone, and preserve specified constraints is a workflow.
Excel presents the harder and potentially more valuable test. “Deeper data analysis” can mean anything from explaining a table to examining patterns across a complicated workbook. The useful version must understand not only visible cell values but formulas, data types, dependencies, missing information, and the business question behind the spreadsheet.
This is also where verification becomes unavoidable. A fluent narrative about a workbook can be dangerously persuasive when the analysis is based on the wrong range, an overlooked filter, a malformed date, or a formula the model misunderstood. GPT-5.6 may reduce the number of prompts required, but it does not eliminate the need to inspect assumptions and calculations.
PowerPoint exposes a different weakness in workplace AI: the gap between generating words and producing a coherent visual argument. CIOL says GPT-5.6 can turn early ideas into more polished presentations, suggesting that Microsoft and OpenAI are targeting structure, formatting, and artifact quality rather than merely slide text.
A useful presentation agent must decide what belongs on each slide, maintain a narrative sequence, distinguish evidence from decoration, and fit content into a template without turning every page into a wall of text. The model also needs to revise the deck when the underlying source changes, rather than treating the presentation as a disposable endpoint.
Across all three applications, the promise is that Copilot will require less micromanagement. The risk is that users may interpret smoother execution as stronger factual reliability.
Better artifact generation changes the form of the review burden; it does not remove it. When AI output arrives as a polished document, analyzed workbook, or executive-ready deck, the mistakes become easier to distribute.

Ultra Mode Turns Copilot’s Model Upgrade Into an Agent Story

GPT-5.6 introduces Ultra mode, a configurable reasoning mode that coordinates multiple AI agents in parallel for complex tasks. RTTNews says the mode is intended to improve computer use, reasoning, and productivity, while CIOL associates it with software engineering, research, and long-running enterprise workflows.
Parallel agents can divide a project into specialized streams. One might inspect source documents, another analyze a spreadsheet, another draft an executive narrative, and another check whether the resulting presentation reflects the evidence.
That approach can shorten elapsed time and broaden the amount of work attempted in one run. It can also increase the system’s operational complexity because the agents must share context, resolve conflicting interpretations, coordinate tools, and combine their outputs into one answer.
The word “parallel” should not be confused with “independent proof.” If several agents inherit the same flawed premise or incomplete source set, they can produce a more elaborate version of the same mistake. Multi-agent execution is most valuable when roles are deliberately separated and the system exposes enough of its plan for a human to challenge it.
For Microsoft 365 Copilot, Ultra mode points toward a future in which a request is no longer one model call. A user might ask Copilot to prepare a quarterly review, and the underlying system could retrieve documents, analyze workbook data, draft the report, create a deck, compare outputs, and revise the final package.
This would be a substantial improvement over moving manually between Word, Excel, PowerPoint, email, and team conversations. It would also mean that a single user instruction can trigger a much larger chain of access and action.
The governance unit must therefore expand from the prompt to the workflow. Administrators need to care about which data was retrieved, which tools were invoked, which actions were permitted, what intermediate artifacts were created, and whether the final output retained a traceable connection to its sources.
Agentic convenience enlarges the blast radius of a bad instruction. The same capability that removes manual coordination can also propagate a mistaken assumption across several files and teams before anyone notices.

Cowork Reveals Microsoft’s Real Product Direction

CIOL says GPT-5.6 strengthens Cowork by helping teams complete cross-functional projects with less manual coordination. That makes Cowork more than another place to chat with a model; it becomes an environment for work that spans people, files, applications, and stages.
Cross-functional projects are exactly where conventional assistants struggle. The required context is fragmented across documents, spreadsheets, presentations, messages, and systems owned by different teams. The job rarely ends with one response because the output of one stage becomes the input to another.
A Cowork workflow could reportedly assemble information, draft materials, analyze data, and help coordinate the resulting deliverables. The practical value is not simply faster content creation but reduced handoff friction between departments.
That places Microsoft in a difficult but advantageous position. It already controls the applications in which much enterprise knowledge work is created, reviewed, and distributed. OpenAI supplies a high-capability model, while Microsoft supplies the identity system, permissions, files, organizational context, application surfaces, and administrative controls.
The partnership works because each company owns something the other would find expensive to reproduce. OpenAI cannot instantly recreate Microsoft’s enterprise distribution and application footprint; Microsoft cannot assume its internal models immediately match OpenAI’s strongest systems on every demanding task.
Yet Cowork also demonstrates why Microsoft cannot remain dependent on one vendor. Once Copilot begins executing end-to-end workflows, model cost, latency, availability, specialization, and contractual control become platform concerns. A diversified model layer is not merely negotiating leverage; it is operational resilience.
OpenAI’s role may consequently become both more important and less exclusive. GPT-5.6 can power the most ambitious parts of the experience while Microsoft uses MAI or other systems for cheaper, specialized, or policy-sensitive steps.

ChatGPT Work Opens a Second Front Inside the Office

Alongside GPT-5.6, OpenAI has launched ChatGPT Work, which CIOL describes as combining ChatGPT with Codex. It can generate presentations, reports, spreadsheets, software, websites, and other deliverables from a single interface.
The product supports browser automation, scheduled tasks, multi-step workflows, tool coordination, project-context maintenance, and the automation of repetitive business processes. OpenAI’s official product material similarly frames Work as a system that gathers context, plans an approach, acts across tools and files, and returns finished artifacts for review.
This creates an unusual competitive geometry. OpenAI is strengthening Microsoft 365 Copilot while simultaneously building an alternative place from which users can perform Office-like work.
ChatGPT Work does not need to replace Word, Excel, or PowerPoint to pressure Microsoft. It only needs to become the preferred starting point for a meaningful share of document, analysis, presentation, and automation workflows.
If the user begins in ChatGPT Work, the Office application may become an output format or final editing destination rather than the primary workspace. That shifts control over user intent and workflow orchestration toward OpenAI.
Microsoft 365 Copilot offers the inverse proposition: remain inside the applications and organizational environment already approved by the business. Microsoft can argue that the safest and least disruptive path to agentic work runs through existing identities, policies, files, and administrative controls.
The competition is therefore not simply GPT-5.6 versus another model. It is Microsoft’s application-centered approach versus OpenAI’s agent-centered approach, even as both companies depend on the same GPT-5.6 family.
CIOL places ChatGPT Work against Anthropic’s Claude Cowork, Google Gemini for Workspace, and Amazon Q Business. The broader contest among Microsoft, Google, Anthropic, Amazon, and OpenAI is moving from chatbot quality toward ownership of the workplace execution layer.
The winning platform will not necessarily be the one that writes the best memo from a blank prompt. It will be the one that can securely collect the right context, complete the largest share of the process, produce editable outputs, recover from errors, and show enough of its work to earn organizational trust.

The Partnership Is Becoming a Supply Chain

OpenAI’s partnership statement emphasizes a shared mission of bringing advanced AI to more people and organizations. The commercial reality underneath that language is becoming more complicated.
Microsoft needs access to frontier models because Copilot’s perceived quality is compared with whatever customers can obtain directly from OpenAI and its competitors. If Microsoft’s suite visibly lags behind standalone AI products, its distribution advantage will not be enough to preserve user enthusiasm.
OpenAI needs Microsoft because enterprise adoption is not just a model-access problem. Businesses require identity management, data controls, application integration, procurement relationships, compliance processes, and support structures. Microsoft already has those channels.
At the same time, each party has a reason to reduce dependence on the other. Microsoft is developing MAI and reportedly inserting it into Word and Excel features. OpenAI is building ChatGPT Work into a direct enterprise workspace rather than remaining only a supplier behind Microsoft’s interface.
This resembles a strategic supply chain more than a conventional product partnership. Microsoft can source capability from OpenAI while developing internal substitutes; OpenAI can use Microsoft’s reach while creating its own route to workplace users.
GPT-5.6 becoming Copilot’s preferred model should therefore not be read as evidence that diversification has stopped. It shows where the quality threshold currently sits.
Microsoft appears willing to replace external models where internal systems are good enough and keep OpenAI where the difference remains commercially significant. OpenAI, in turn, is making efficiency a core part of GPT-5.6’s pitch so that its frontier capability is harder to remove on cost grounds.

The Three-Model Family Gives Microsoft More Levers

Sol, Terra, and Luna offer Microsoft a more nuanced procurement and routing strategy than a single flagship model would. Instead of deciding whether to use GPT-5.6, Microsoft can decide which GPT-5.6 variant is appropriate for each workload.
Sol is positioned for the tasks where failure, ambiguity, or long execution makes advanced reasoning worthwhile. That includes difficult coding, cybersecurity analysis, scientific work, and complex professional workflows.
Terra is the natural candidate for broad productivity use where Microsoft wants strong results without paying the full cost of the flagship tier. It may be the most strategically important member of the family because balanced models typically encounter the widest range of enterprise tasks.
Luna targets high-volume, cost-sensitive work. Its existence makes OpenAI more competitive with Microsoft’s internal-model strategy because OpenAI can offer a lower-cost route for tasks that do not require Sol’s reasoning ceiling.
The family also lets Microsoft compare internal and external models at several performance levels. MAI does not have to defeat Sol to create value; it only needs to outperform the external alternative on cost, latency, control, or specialization for a defined workload.
For users, the benefit of this competition should be a Copilot that wastes less time and compute. The drawback is reduced transparency if the product does not clearly communicate which model, mode, or policy produced an important result.
A model label is not mere enthusiast trivia in an enterprise setting. Different models can have different behavior, tool support, reliability profiles, data-processing arrangements, and suitability for regulated work.
Administrators do not necessarily need users selecting a model for every request. They do need enough visibility to test, govern, and explain the system they have deployed.

Admins Must Test Workflows, Not Model Marketing

A phased rollout across enterprise and commercial subscription plans means availability will not necessarily arrive for every tenant, application, or user at the same time. Organizations should resist treating the launch announcement as evidence that their production environment has already changed uniformly.
The first task is inventory. Administrators need to determine where GPT-5.6 appears, whether users can select a variant or reasoning mode, which applications route automatically, and whether Cowork behavior differs from Word, Excel, PowerPoint, or Copilot Chat.
The second task is evaluation using the organization’s actual work. Generic prompts will not expose problems involving complex templates, formulas, permissions, terminology, retention requirements, or cross-functional approval chains.
The third task is governance. Ultra mode and other agentic workflows may use more tools and data than a simple chat request. Organizations need a clear distinction between generating a suggestion and taking an action.

Action checklist for admins

Confirm GPT-5.6 availability separately in Word, Excel, PowerPoint, Copilot Chat, and Cowork rather than assuming tenant-wide parity.
Document whether Sol, Terra, Luna, or Ultra mode can be selected by users or is chosen automatically.
Re-run approved evaluation sets for document drafting, workbook analysis, presentation creation, and cross-functional workflows.
Test permissions with overshared, restricted, and intentionally inaccessible content before enabling broader agentic use.
Require human review for financial analysis, security findings, external communications, and consequential automated actions.
Monitor output quality, latency, prompt count, tool use, and correction rates so the rollout is measured by completed work rather than model branding.

Organizations should also prepare users for behavioral changes. If a familiar Copilot task suddenly produces a more ambitious output, employees may assume the system has become categorically reliable rather than incrementally more capable.
Training should focus on review techniques: checking the source set, verifying spreadsheet ranges and formulas, inspecting document claims, and testing whether a presentation’s conclusions follow from the evidence. A polished artifact should be treated as a draft with a shorter path to completion, not an automatic transfer of accountability.

Programmatic Tool Calling Raises the Stakes for Developers

CIOL reports that GPT-5.6 adds programmatic tool calling for developers. This allows AI applications to coordinate tools and manage intermediate workflow steps with less manual intervention.
For developers, that can reduce the amount of rigid orchestration code needed to build multi-stage applications. A model can decide when to retrieve information, invoke a service, transform an intermediate result, or request another tool.
The danger is that flexible orchestration can hide failure in the middle of a successful-looking workflow. A tool may return incomplete data, an agent may select the wrong function, or an intermediate result may be misinterpreted before the final response is produced.
Production implementations need explicit boundaries. Tool permissions should be narrow, side effects should be distinguishable from read-only operations, and consequential actions should require approval or a reversible execution path.
Logging must capture more than the final answer. Developers need records of which tools were called, what arguments were supplied, what each tool returned, and how the system handled errors or conflicting results.
Programmatic tool calling also intensifies the importance of predictable model behavior. A workflow that works reliably with one variant may behave differently when routed to another. If Microsoft or an application automatically changes the underlying model, regression testing becomes a continuing operational requirement.
The goal should not be to prevent model routing. It should be to ensure that routing does not silently invalidate the assumptions on which a business process was approved.

What Windows and Microsoft 365 Users Will Actually Notice

For ordinary users, the launch will probably not arrive as a dramatic new application. GPT-5.6 is being inserted into tools they already use, and its success will be measured by how often it avoids making itself the subject of the work.
A Word user may notice that a document requires fewer instructions to reach the intended structure. An Excel user may receive a more useful analysis of a complicated workbook. A PowerPoint user may get a presentation that looks closer to a deliverable and less like paragraphs copied onto slides.
Copilot Chat may become better at carrying a request across several steps, while Cowork may handle projects that previously required repeated manual coordination. Users may also see different capabilities at different times because the rollout is expanding in phases.
They should not assume every surface is backed by the same variant or routing policy. Word and Excel may contain a mixture of GPT-5.6, Microsoft’s reportedly deployed MAI models, and application-specific systems.
That mixture is not automatically a problem. Most users do not care which model rewrites a sentence if the result is fast, accurate, and appropriately governed.
They will care when identical-looking Copilot experiences behave inconsistently, when a feature disappears or moves, or when an important result cannot be reproduced. Microsoft’s challenge is to hide irrelevant infrastructure complexity without hiding information customers need to trust the system.
Nitin Agrawal, President, Copilot & Agents Core at Microsoft, said, “We can't wait for customers to see what GPT-5.6 in Microsoft 365 will do.” The decisive test is whether customers see a dependable improvement in their work rather than merely a new model name in the interface.

The Copilot Upgrade in Six Practical Conclusions

GPT-5.6 is a significant Microsoft 365 development because it joins frontier reasoning, lower-cost model variants, and parallel-agent execution inside a platform already moving from assistance toward delegated work. The announcement also confirms that Microsoft’s diversification and its OpenAI partnership are occurring at the same time.

GPT-5.6 is Copilot’s preferred model, but that does not establish that it handles every request.
Sol, Terra, and Luna give Microsoft distinct capability-and-cost tiers for different workloads.
Microsoft’s reported MAI deployment in Word and Excel is consistent with selective routing rather than a full OpenAI replacement.
Ultra mode makes governance of tools, permissions, and intermediate actions as important as review of the final answer.
ChatGPT Work turns OpenAI into both Microsoft’s model supplier and a direct competitor for workplace workflow control.
The most useful adoption metric is fewer corrections and more verified work completed, not benchmark scores or model visibility.

The GPT-5.6 announcement does not restore a simple Microsoft-OpenAI alliance because that simple alliance no longer exists. It reveals a more durable arrangement in which Microsoft builds the routing, governance, application, and internal-model layers while OpenAI competes to remain the premium intelligence inside them—and as both companies push deeper into agentic work, customers will have to demand that the convenience of invisible orchestration is matched by visible control.

References

Primary source: RTTNews
Published: 2026-07-10T21:12:07.776781

OpenAI Says GPT-5.6 Will Be Preferred AI Model For Microsoft 365 Copilot

OpenAI says its newly launched GPT-5.6 will become the preferred model powering Microsoft 365 Copilot, reinforcing its partnership with Microsoft amid speculation that the two companies are drifting apart.

www.rttnews.com
Independent coverage: ciol.com
Published: 2026-07-10T13:12:07.781973

OpenAI Brings GPT-5.6 to Microsoft 365 Copilot, Launches ChatGPT Work

OpenAI launches GPT-5.6 and ChatGPT Work, introducing AI-powered enterprise workflows, coding, research, and document creation in a unified workspace.

www.ciol.com
Official source: help.openai.com

Creating and editing documents, spreadsheets, and presentations with ChatGPT Work | OpenAI Help Center

Create editable files from instructions, source material, or an existing template.

help.openai.com
Official source: openai.com

ChatGPT Work with GPT-5.6 | OpenAI

ChatGPT Work, powered by GPT-5.6, helps teams take on ambitious work and turn goals into finished outputs. Connect tools, automate tasks, and keep projects moving.

openai.com
Related coverage: chatgpt.com

https://chatgpt.com/codex
Official source: academy.openai.com

ChatGPT Work Resource Guide - Resource | OpenAI Academy

Unlock the new opportunities of the AI era by equipping yourself with the knowledge and skills to harness artificial intelligence effectively.

academy.openai.com

Official source: edunewsletter.openai.com

Introducing ChatGPT Work. A More Capable ChatGPT for Education

Introducing ChatGPT Work, a new way for faculty and staff to carry out longer, multi-step work across approved files, connected apps, and the web.

edunewsletter.openai.com
Related coverage: thesheffieldpress.com

OpenAI launches ChatGPT Work to automate multi-step workplace tasks | The Sheffield Press

OpenAI rolled out ChatGPT Work, a GPT-5.6 agent that can run tasks across apps and files. The new desktop app and Sites beta extend it beyond chat.

thesheffieldpress.com
Related coverage: qualeai.it

ChatGPT Work 2026: quando usarlo | QualeAI

ChatGPT Work, Chat e Codex a confronto: cosa cambia nella nuova app ChatGPT, piani, limiti e quando usare ogni modalità.

www.qualeai.it
Official source: deploymentsafety.openai.com

GPT-5.6 Preview System Card

PDF document

deploymentsafety.openai.com
Related coverage: tech.yahoo.com

OpenAI says GPT 5.6 is the ‘preferred model’ for Microsoft Copilot 365 amid breakup chatter

OpenAI's new family of models will continue to power Microsoft's suite of workplace and productivity apps.

tech.yahoo.com
Official source: learn.microsoft.com

Release Notes for Microsoft 365 Copilot | Microsoft Learn

Lists the features that reach General Availability in each release of Microsoft 365 Copilot.

learn.microsoft.com
Related coverage: marketscreener.com

https://www.marketscreener.com/news/openai-gets-us-approval-for-broad-gpt-5-6-rollout-axios-reports-ce7f5ed8df81f020
Related coverage: github.blog

OpenAI's GPT-5.6 Sol, Terra, and Luna are now available in GitHub Copilot - GitHub Changelog

OpenAI’s GPT-5.6 family is now rolling out in GitHub Copilot. GPT-5.6 comes in three variants, Sol, Terra, and Luna, so you can match the model to the job, whether that’s…

github.blog
Related coverage: newsbytesapp.com

GPT 5.6 is the 'preferred model' for Microsoft 365 Copilot

OpenAI reaffirmed GPT 5.6 as the top choice for Microsoft Copilot 365, following reports of Microsoft integrating its in-house MAI models to reduce operational costs.

www.newsbytesapp.com
Related coverage: sahmcapital.com

Microsoft unveils AI upgrades, rolls out Copilot Cowork to early-access customers

By Deborah Mary Sophia March 30 (Reuters) - Microsoft MSFT.O on Monday unveiled new features in its Copilot research assistant that would allow users to utilize multiple AI models simultaneously within the same workflow, the latest move by the tec...

www.sahmcapital.com
Official source: news.microsoft.com

Copilot Cowork Now Available in the Frontier Program - Source EMEA

news.microsoft.com
Related coverage: investing.com

OpenAI launches GPT-5.6 model family with Sol flagship By Investing.com

OpenAI launches GPT-5.6 model family with Sol flagship

www.investing.com
Official source: techcommunity.microsoft.com

GA announcement

PDF document

techcommunity.microsoft.com
Related coverage: intuitionlabs.ai

microsoft copilot pricing licensing guide for business

PDF document

intuitionlabs.ai
Related coverage: techriver.com

Microsoft Word - Microsoft O365 December 2025 Updates.docx

PDF document

www.techriver.com
Official source: cdn-dynmedia-1.microsoft.com

Slide 1

PDF document

cdn-dynmedia-1.microsoft.com
Official source: microsoft.com

Microsoft 365 Roadmap | Microsoft 365

The Microsoft 365 Roadmap lists updates that are currently planned for applicable subscribers. Check here for more information on the status of new features and updates.

www.microsoft.com
Official source: microsoftpartners.microsoft.com

Enabling Frontier Workers

ChatGPT · 2026-07-10T22:18:23-0400

Microsoft 365 Copilot began rolling out GPT-5.6 as its preferred and default model on July 9 across Word, Excel, PowerPoint, Copilot Chat, and Cowork, targeting complex, multi-step work while automatically routing requests between models according to Microsoft’s optimization criteria. The day-zero deployment matched GPT-5.6’s general availability in ChatGPT and Codex, turning a model launch into an immediate enterprise-software event. But the most consequential change is not simply a smarter Copilot. It is Microsoft’s decision to make model selection an increasingly invisible infrastructure function inside the productivity suite.
That distinction matters because “preferred,” “default,” and “used for every interaction” are not interchangeable claims. Crypto Briefing described GPT-5.6 as the default brain powering Microsoft 365 Copilot, while TechCrunch emphasized that the preferred-model designation did not establish that every prompt would reach GPT-5.6. The apparent contradiction is resolved by the rollout mechanics: GPT-5.6 can be the favored destination for demanding work while Microsoft’s routing layer continues sending other requests elsewhere.

Microsoft Is Upgrading the Office Workflow, Not Just the Chatbot

Model launches are usually discussed as contests of benchmark scores, context windows, and prompt demonstrations. Microsoft’s deployment changes the unit of competition. The question is no longer whether GPT-5.6 can answer a difficult prompt in isolation, but whether Microsoft can insert it into an operating workflow that begins with company data, crosses several applications, and ends with a usable business artifact.
That is why the stated focus on “complex, multi-step tasks” is more important than the model number. Microsoft and OpenAI are positioning the upgrade for work such as comparing options in Excel, building structured planning documents in Word, running troubleshooting workflows, and performing data analysis that would otherwise require a user to divide the problem into several prompts and manually reconcile the answers.
A basic writing assistant produces a paragraph. A workflow assistant must understand the assignment, gather the relevant context, preserve constraints, organize intermediate results, and produce something that survives contact with the next person in the process. In Office, that final requirement is unforgiving: an analysis must fit the workbook, a plan must fit the document, and a presentation must fit the organization’s actual communication practices.
The integration across Word, Excel, PowerPoint, Copilot Chat, and Cowork consequently represents more than broad product coverage. It gives Microsoft multiple places to intercept knowledge work: the document being written, the spreadsheet being analyzed, the presentation being assembled, the conversation in which a task is formulated, and the agentic workspace in which actions can be coordinated.

Microsoft 365 surface	GPT-5.6’s intended role	Representative task	Practical consequence
Word	Structure and synthesize longer work	Planning documents	Less manual assembly of sections, constraints, and supporting material
Excel	Reason across data and alternatives	Option comparisons and data analysis	More emphasis on interpretation, not merely formulas or summaries
PowerPoint	Turn analysis into an organized artifact	Multi-step presentation work	A shorter path from source material to a usable narrative
Copilot Chat	Coordinate complex requests conversationally	Troubleshooting workflows	The conversation becomes an entry point into a broader work process
Cowork	Carry work across Microsoft 365	Multi-application tasks	Copilot moves closer to execution rather than recommendation

For users, the immediate benefit should be fewer seams. The old failure mode of productivity AI was not always that it could not complete any individual step; it was that the user remained responsible for decomposition, verification, transfer, and reformatting between steps. Microsoft is betting that GPT-5.6 can absorb more of that connective labor.
For IT departments, however, fewer visible seams can mean more hidden dependencies. Once a model participates in a chain of document retrieval, reasoning, transformation, and action, diagnosing a bad result becomes harder than reviewing a single chatbot response. The model is only one component of a system that also includes permissions, grounding, application state, routing rules, and Microsoft’s surrounding orchestration.

Day-Zero Deployment Turns Model Releases Into Enterprise Change Events

The July 9 integration was described as a day-zero deployment because it coincided with GPT-5.6’s general availability in ChatGPT and Codex. That synchronization is strategically significant. Microsoft did not wait for the model to become old news, observe it from a distance, and introduce it months later as an optional Office enhancement.
Instead, the company treated the release of a frontier model as something Microsoft 365 should consume immediately. That shortens the distance between model research and enterprise production, but it also compresses the traditional period in which customers, security teams, and software vendors learn how a new model behaves.
A day-zero rollout does not mean every tenant receives identical behavior at the same moment. Crypto Briefing describes a phased approach in which Microsoft is monitoring performance telemetry before pushing GPT-5.6 into every Copilot interaction globally. That qualification is essential: the launch date marks the start of operational deployment, not proof of uniform worldwide saturation.
The phased approach gives Microsoft room to manage capacity, reliability, output quality, and cost. It also lets the company compare how GPT-5.6 performs inside real Microsoft 365 workloads rather than assuming results from a general model evaluation will transfer cleanly into heavily permissioned corporate environments.
This is the new cadence enterprise customers must prepare for. Historically, an Office upgrade arrived as a feature release that could be documented, tested, and communicated. In an AI-mediated service, Microsoft can alter the effective capability of Word or Excel by changing the model, the router, the surrounding instructions, or the thresholds that determine which requests qualify for deeper reasoning.
The visible interface may barely move while the service underneath it changes materially. That makes AI releases resemble cloud-service modifications more than desktop application upgrades: centrally operated, incrementally deployed, telemetry-driven, and capable of changing results without a conventional software installation.
Organizations that still treat Copilot as a static add-on will struggle with this model. Its behavior is increasingly the product of a living service stack, and that stack can evolve faster than internal validation procedures, training material, or approved-use documentation.

The Router, Not GPT-5.6, Is Becoming Copilot’s Real Control Plane

Microsoft is using automatic task routing to decide which model handles a request according to optimization criteria. Users can manually select GPT-5.6 where a model picker is available, but algorithmic matching remains the default behavior described in the rollout.
The appeal is obvious. Most employees do not want to study model catalogs before drafting a memo or analyzing a worksheet. Microsoft can reserve a more capable model for harder tasks, use faster or less expensive resources for routine requests, and respond to capacity conditions without forcing users to make an infrastructure decision every time they open Copilot.
The router can also become a substantial economic advantage. Early feedback cited by Crypto Briefing points to improved performance per token, meaning more useful work may be completed for a given amount of model computation. If Microsoft can combine that efficiency with accurate routing, it can improve user-visible capability without paying the maximum inference cost for every prompt.
But preferred does not mean exclusive. TechCrunch’s coverage correctly focused on the ambiguity of the term, noting that the announcement did not prove every Microsoft 365 request would be served by GPT-5.6. Automatic routing makes the distinction operational rather than semantic: GPT-5.6 may be the favored model for demanding work while other requests continue flowing to different systems.
That means two users can issue superficially similar requests and encounter different performance characteristics. A prompt may be interpreted as routine in one context and complex in another. Service availability, application, tenant configuration, available controls, and Microsoft’s optimization criteria may all influence what happens after the user presses Enter.
From Microsoft’s perspective, the router is an abstraction layer. It allows the company to improve or replace underlying components while preserving the Copilot interface. From an enterprise perspective, the same abstraction can obscure exactly which technology processed sensitive information or produced a consequential recommendation.
This is not merely a user-interface complaint. It affects reproducibility, validation, incident investigation, and records management. If an employee cannot determine which model handled a request, an administrator may have difficulty explaining why a previously reliable prompt changed, whether a documented test still applies, or which processing commitments governed a particular workflow.
Microsoft’s broader documentation already distinguishes among models hosted and operated by Microsoft, third-party models acting as subprocessors, and independent processors operating under their own terms. Those categories can carry different contractual, privacy, and administrative implications. The routing layer therefore sits directly between convenience and governance.
The ideal enterprise router would not require employees to become model engineers, but it would preserve enough metadata for authorized administrators to reconstruct important events. The system should be simple at the point of use and detailed at the point of audit. Those goals are compatible, but only if Microsoft designs for both.

Better Performance per Token Changes Copilot’s Economics

“Performance per token” sounds like a narrow infrastructure metric, but it is central to the economics of enterprise AI. Every long document, workbook analysis, troubleshooting chain, and cross-application workflow consumes compute. If the model needs fewer tokens, fewer retries, or less user correction to produce a useful result, the cost of deploying AI across a large organization changes substantially.
The crucial unit is not the cheapest response. It is the cost of an accepted result. A model that produces a quick but incomplete analysis can be more expensive than a slower model if an employee must repair the output, repeat the prompt, or redo the work manually.
This is where GPT-5.6’s focus on complex tasks could matter most. Multi-step work amplifies small inefficiencies. A weak intermediate conclusion can contaminate the remaining workflow, while an unnecessary detour can generate additional tokens, tool calls, and human review.
Improved efficiency can therefore benefit both sides of Microsoft’s platform. Customers may receive more capable behavior without corresponding increases in waiting time or service limits, while Microsoft can reduce the infrastructure cost of providing Copilot at enterprise scale.
The complication is that automatic routing makes those economics difficult for customers to inspect. An organization may observe that Copilot is faster or better on average without knowing whether the gain came from GPT-5.6, a routing change, a revised system instruction, or a different allocation of compute.
For procurement and governance teams, that makes outcome-based evaluation more useful than model-brand evaluation. The operational questions are whether a workflow succeeds, how often it requires correction, how long it takes, what data it touches, and whether the result can be audited. The model label matters, but it is not a substitute for measuring the service.
Microsoft has every incentive to make the model layer fungible. It wants Copilot to select the best available resource rather than bind the suite permanently to one model. Enterprise customers have the opposite incentive in high-consequence workflows: they need enough stability and traceability to know that yesterday’s validation still means something today.
The resulting tension will define the next phase of Copilot administration. Microsoft will optimize a fleet; customers will try to certify a process.

The OpenAI Partnership Is Deepening and Becoming Less Exclusive at the Same Time

Microsoft and OpenAI’s partnership dates to 2019, and the day-zero GPT-5.6 integration demonstrates that the relationship can still move frontier technology into Microsoft’s products with exceptional speed. The deployment reaches a productivity suite used by Fortune 500 companies, giving OpenAI immediate access to business workflows that standalone chatbot vendors must often approach through connectors or separate applications.
Yet the announcement arrived against reporting that Microsoft was also relying more heavily on its own models in workplace applications to reduce costs. Bloomberg reported that Microsoft had begun using internally developed alternatives for some prompts, and TechCrunch observed that the GPT-5.6 preferred-model announcement did not invalidate that account.
There is no need to force these developments into a breakup-or-reconciliation narrative. They are two sides of the same platform strategy. Microsoft can maintain privileged access to OpenAI’s frontier capabilities while building an internal model portfolio and routing simpler work toward whichever resource offers an acceptable result at lower cost.
The partnership is therefore compounding at the application layer while loosening at the infrastructure layer. OpenAI remains strategically important because frontier models can raise the ceiling of Copilot. Microsoft’s own systems remain strategically important because they can lower the floor of what the company must pay to serve routine requests.
This is why the term “preferred” does useful work for both companies. OpenAI can point to prominent placement across Microsoft 365. Microsoft can endorse GPT-5.6 without promising that every prompt, tenant, application, or workload will depend exclusively on it.
The real competition is no longer only between model developers. It is between orchestration platforms capable of combining models, enterprise data, permissions, applications, and telemetry into a service users can trust. Microsoft’s advantage is that it controls the productivity surface in which the work is already taking place.
OpenAI’s advantage is that Microsoft still considers its latest model important enough for day-zero integration. Each company needs the other, but neither wants its future reduced to the other’s roadmap.

The Trust Gap Is Now About Observability

Enterprise users have already raised concerns about not always knowing which model handles a request. Crypto Briefing identifies model transparency alongside data sovereignty and the treatment of proprietary information as a central customer concern.
These questions should not be collapsed into a generic fear that “the AI has company data.” They are separate governance issues. Data sovereignty concerns where processing occurs and which commitments apply. Model transparency concerns which system performed the processing. Training concerns whether customer information is used to improve future models.
Microsoft’s documentation states that some models used in its online services are hosted on Azure and provided directly by Microsoft under Microsoft’s enterprise safeguards. It also documents separate classifications for third-party AI subprocessors and independent processors. In a multi-model environment, admins need to understand which category is available, enabled, or eligible for routing in each workload.
The supplied reporting says enterprise customers are asking whether proprietary information is used to train future models. The responsible administrative response is not to infer an answer from the presence of an OpenAI model name. Organizations should rely on the contractual terms, data-protection documentation, tenant settings, and product-specific disclosures governing the Microsoft 365 service they actually use.
Model branding can mislead in both directions. A familiar model name does not necessarily mean data is sent through the same consumer service, under the same terms, or with the same handling as a public chatbot. Conversely, a Microsoft interface does not eliminate the need to identify any third party participating in processing.
Routing is now policy, even when it is presented as a performance feature. A routing decision can determine not only quality and latency but also which provider, hosting arrangement, regional boundary, and administrative control becomes relevant.
The missing enterprise feature is a durable routing receipt: metadata showing which model class handled a request, when it was processed, what policy allowed that selection, and whether fallback occurred. Employees may not need to see all of that in every response, but compliance teams, security investigators, and service owners will increasingly need it.
Without such observability, organizations must validate a black box whose internal path can change. That is manageable for low-risk drafting. It is much less comfortable for financial analysis, regulated communications, troubleshooting procedures, or decisions that must later be defended.

The Rollout Requires Workflow Testing, Not Model Tourism

The temptation after a model upgrade is to open the picker, select GPT-5.6, and try a collection of clever prompts. That can reveal obvious changes, but it is not an enterprise test plan.
Admins should begin with recurring business workflows that already have known inputs and expected outputs. A useful evaluation set might include a standard option-comparison workbook, a planning-document template, a troubleshooting procedure, and a data-analysis task with independently verified conclusions.
The goal is not to prove that GPT-5.6 is universally superior. It is to determine where the new service produces an acceptable result with less human effort, where routing introduces variability, and where users must continue applying formal review.
Testing should also distinguish manual model selection from automatic routing. A workflow that succeeds when GPT-5.6 is explicitly chosen may behave differently when Copilot decides which model to use. Both paths matter because most users will accept the default rather than manage the picker.
Organizations should record prompts, source files, output quality, completion time, required corrections, and any visible model information. That creates a baseline for future changes and helps separate genuine service regressions from altered source data or permissions.
The phased rollout makes this especially important. Different users or tenants may encounter the upgrade at different points, and Microsoft is reportedly using telemetry before extending GPT-5.6 to every Copilot interaction globally. Internal support teams should avoid assuming that every user reporting different behavior is mistaken.

Action checklist for admins

Inventory the Word, Excel, PowerPoint, Copilot Chat, and Cowork workflows that involve sensitive data or consequential decisions.
Test representative multi-step tasks under both automatic routing and manual GPT-5.6 selection where the picker is available.
Capture baseline outputs, completion times, correction rates, and any visible model or routing information.
Review Microsoft’s current data-processing, data-sovereignty, subprocessor, and enterprise-protection documentation for enabled Copilot experiences.
Tell users that “preferred” does not guarantee GPT-5.6 handles every request.
Require human verification for financial, legal, security, operational, or externally published outputs.
Monitor tenant communications and performance reports throughout the phased rollout rather than treating July 9 as universal completion.

The most useful internal guidance will be task-specific. “Check AI output” is too vague to change behavior. “Verify every calculated comparison against the workbook’s source range before circulating the recommendation” is an enforceable control.

Copilot Is Moving From Assistance Toward Delegation

The inclusion of Cowork clarifies Microsoft’s direction. Word, Excel, and PowerPoint place AI inside individual artifacts; Copilot Chat gives users a conversational control surface; Cowork moves toward carrying out coordinated work across Microsoft 365.
As that progression continues, the risk model changes. A drafting assistant can create inaccurate text. A delegated workflow can create inaccurate text, place it in the wrong artifact, carry a mistaken assumption into another step, or prepare an action based on it.
The remedy is not to freeze Copilot at simple autocomplete. It is to design approval points around the consequences of the task. Low-risk internal formatting may need little intervention, while external communication, operational changes, or decisions based on sensitive analysis require explicit review.
GPT-5.6’s intended strength in multi-step work may make these controls more important, not less. Better models encourage users to delegate larger assignments. As the success rate rises, attention can fall, and intermittent failures become harder to notice because the surrounding work looks polished.
Microsoft’s strongest product opportunity is to make review part of the workflow rather than an afterthought. Copilot should expose assumptions, sources, transformations, and pending actions at the point where a human can still correct them. The enterprise value is not autonomous output at any cost; it is controlled delegation with less friction.
This also reframes user training. Employees do not need a course on every available model. They need to know how to define a task, constrain access, recognize uncertainty, inspect evidence, and approve consequential actions.
Model literacy still matters, particularly when a picker is available. But process literacy will matter more. The user must understand what Copilot was allowed to do, what it actually did, and which parts remain their responsibility.

The Crypto-Free Announcement Reveals Where the Infrastructure Race Is Actually Happening

Crypto Briefing noted that the announcement contained no references to crypto, blockchain, or digital assets. That absence is informative because it strips away a familiar investment narrative and leaves the more consequential infrastructure contest in view.
Microsoft and OpenAI are competing for the economics of ordinary business work: documents, spreadsheets, presentations, troubleshooting, planning, and analysis. These tasks are individually mundane but collectively enormous, repeated across organizations every working day.
The winner will not be determined solely by who releases the most impressive model. It will be determined by who can deliver reliable intelligence inside existing workflows at a cost that survives mass deployment.
Microsoft’s automatic routing is an answer to that economic problem. It can use stronger models where their additional capability is valuable and optimize the remainder of the workload according to cost, speed, availability, and policy. The router becomes the mechanism through which expensive frontier intelligence is rationed.
That makes the AI infrastructure race more capital-intensive and more operationally complex. Every improvement in model efficiency invites larger workloads. Every successful workflow increases demand for inference, telemetry, storage, governance, and regional capacity.
For customers, the danger is assuming that an apparently simple Copilot request maps to a simple service transaction. Behind the interface may be retrieval, policy checks, model routing, generation, application-specific processing, and further orchestration. The convenience is real, but so is the infrastructure required to produce it.
The absence of crypto language therefore does not make the announcement financially uninteresting. It makes the economic thesis more concrete: Microsoft and OpenAI are attempting to turn model intelligence into a utility embedded in the software through which corporate work already flows.

What July 9 Changes for Microsoft 365 Teams

The practical reading of the rollout is neither “every Copilot prompt now uses GPT-5.6” nor “this is merely another optional model.” GPT-5.6 is positioned as the preferred/default engine for harder work, but Microsoft’s router and phased deployment remain central to the experience.

GPT-5.6 went live in Microsoft 365 Copilot on July 9 alongside its general availability in ChatGPT and Codex.
The rollout covers Word, Excel, PowerPoint, Copilot Chat, and Cowork.
Microsoft is targeting complex, multi-step tasks rather than only improving short-form chat.
Automatic task routing means GPT-5.6 may be preferred without handling every request.
Manual selection is available only where the product exposes a model picker.
Admins should evaluate workflows, data handling, and routing transparency—not just model output quality.

The upgrade raises Copilot’s potential ceiling, particularly for work that previously forced users to connect several AI-generated steps by hand. It also raises the standard Microsoft must meet for transparency, because a productivity platform cannot ask enterprises to delegate more work while revealing less about how that work was processed.
GPT-5.6’s arrival in Microsoft 365 is ultimately a test of whether Microsoft can make frontier models feel routine without making their operation inscrutable. If the company can pair better performance per token with visible controls, durable audit information, and predictable enterprise protections, July 9 will look like the moment Copilot matured from an embedded chatbot into a true workflow layer. If it cannot, the model may become more capable while the service around it becomes harder to trust—and the next phase of the AI race will be decided not by intelligence alone, but by which platform can finally make intelligent routing accountable.

References

Primary source: Crypto Briefing
Published: 2026-07-10T15:12:08.515892

Microsoft 365 Copilot adopts GPT-5.6 as preferred model, and the AI infrastructure race just got more expensive

Microsoft 365 Copilot now runs GPT-5.6 as its default model across Word, Excel, and PowerPoint, deepening the OpenAI partnership with day-zero enterprise

cryptobriefing.com
Official source: support.microsoft.com

Standard versus priority access to features in Microsoft 365 Copilot Chat | Microsoft Support

Learn how Copilot Chat is able to understand your prompts.

support.microsoft.com
Official source: learn.microsoft.com

Release Notes for Microsoft 365 Copilot | Microsoft Learn

Lists the features that reach General Availability in each release of Microsoft 365 Copilot.

learn.microsoft.com
Official source: microsoft.com

Available today: GPT-5 in Microsoft 365 Copilot | Microsoft 365 Blog

Microsoft has launched GPT-5 in Microsoft 365 Copilot and Copilot Studio, featuring a system for more efficient problem-solving.

www.microsoft.com
Official source: marketingassets.microsoft.com

Microsoft 365 Copilot Experience Guide

PDF document

marketingassets.microsoft.com
Related coverage: windowscentral.com

Microsoft Drops GPT-5 Into Copilot—No Subscription Needed | Windows Central

OpenAI's much-hyped GPT-5 AI model is now available as part of Microsoft Copilot. Here's what you need to know.

www.windowscentral.com

Related coverage: techradar.com

Microsoft makes Copilot Cowork open to everyone, and wants to help you tackle even the trickiest work tasks | TechRadar

Copilot Cowork gets an upgrade as it opens to all users

www.techradar.com
Related coverage: axios.com

Microsoft puts a price tag on its AI "copilots" for business

$30 a month per user could add up to billions for the software giant.

www.axios.com
Related coverage: itpro.com

Everything you need to know about the new E7 Microsoft 365 tier, including features, pricing, and release date | IT Pro

The new premium bundle for Microsoft 365 adds AI capabilities to traditional tiers

www.itpro.com
Official source: blogs.microsoft.com

Introducing the First Frontier Suite built on Intelligence + Trust - The Official Microsoft Blog

Today Microsoft is announcing: Wave 3 of Microsoft 365 Copilot Expanded model diversity with Claude and next-gen OpenAI models available today General availability of Agent 365 on May 1 for $15 per user General availability of the new Microsoft 365 E7: The Frontier Suite on May 1 for $99 per...

blogs.microsoft.com
Official source: developer.microsoft.com

Not all model upgrades are upgrades - Microsoft for Developers

A new model drops with lower per-token pricing and better benchmarks. You switch. A week later someone asks why the agent is burning 12x more tokens on

developer.microsoft.com
Related coverage: newsbytesapp.com

GPT 5.6 is the 'preferred model' for Microsoft 365 Copilot

OpenAI reaffirmed GPT 5.6 as the top choice for Microsoft Copilot 365, following reports of Microsoft integrating its in-house MAI models to reduce operational costs.

www.newsbytesapp.com
Related coverage: tomsguide.com

Biggest Microsoft Build 2026 announcements — agentic AI, RTX Spark Dev Box, GitHub Copilot app, new MAI models, and more | Tom's Guide

All the big news from Microsoft's AI-focused event

www.tomsguide.com

ChatGPT · 2026-07-11T12:46:45-0400

OpenAI announced GPT-5.6 for Microsoft 365 Copilot on July 10, 2026, positioning its newest model family across Word, Excel, PowerPoint, Copilot Chat, and Cowork. The verified product promises are focused: fewer prompts in Word, deeper and more token-efficient analysis in Excel, higher-quality PowerPoint output, and more capable multi-step workflows in Copilot Chat and Cowork.
The important word is not merely newest, but preferred. OpenAI CEO Sam Altman said GPT-5.6 is now Microsoft 365 Copilot’s preferred model. That gives the release strategic significance, but it does not establish that every Copilot request, user, tenant, or application will use GPT-5.6 in the same way.

What changed

GPT-5.6 has been identified as the preferred model in Microsoft 365 Copilot.

The named Microsoft 365 surfaces are Word, Excel, PowerPoint, Copilot Chat, and Cowork.

The promised improvements include fewer prompts in Word, deeper and more token-efficient Excel analysis, higher-quality PowerPoint output, and more capable Copilot Chat and Cowork workflows.

What has not been confirmed

The supplied information does not confirm per-tenant availability or rollout timing.

It does not establish whether users or administrators can select a particular GPT-5.6 model.

It does not identify required license SKUs, regional availability, tenant settings, administrative controls, or exact enablement procedures.

It does not provide a verified Microsoft 365 Admin Center path or configuration name.

This article should therefore be read as deployment analysis and an evaluation framework, not as a step-by-step administration guide.

The commercial thesis is straightforward. If GPT-5.6 can produce a polished document, interrogate a workbook, assemble a presentation, and support a long-running workflow with less prompting and more efficient use of context, Copilot becomes easier to evaluate as a work platform rather than merely a drafting assistant.
The risk is equally straightforward. As AI performs more of the work between a user’s request and a finished artifact, organizations need stronger methods for reviewing sources, testing assumptions, measuring corrections, and approving consequential actions. A more capable model does not make those controls obsolete. It makes them more important.

“Preferred” Is a Strategic Word, Not a Routing Guarantee

Altman’s statement that GPT-5.6 is the preferred model in Microsoft 365 Copilot confirms its importance within the product. It should not, however, be read as proof that every Copilot request will immediately and exclusively run through the same GPT-5.6 model or configuration.
The supplied information does not describe Microsoft’s model-selection logic. It does not say whether Copilot chooses models by application, task complexity, cost, geography, compliance requirements, availability, or customer configuration. Those may be plausible design possibilities for an enterprise AI service, but they are not confirmed product behavior in the available facts.
“Preferred” should consequently be interpreted narrowly. GPT-5.6 has a favored position in Microsoft 365 Copilot, but the announcement does not explain how that preference is implemented for an individual request. It also does not confirm whether users will see a model selector, whether administrators can enforce a model, or whether the model will be available uniformly across organizations.
That distinction matters when evaluating the release. The consequential question is not whether “GPT-5.6” appears beside a prompt box. It is whether the promised improvements become visible in actual work:

Does Word require fewer rounds of prompting to produce an acceptable document?
Does Excel perform deeper analysis without wasting context or obscuring assumptions?
Does PowerPoint produce a more coherent and presentation-ready result?
Do Copilot Chat and Cowork complete complex workflows more successfully?
Can users verify the sources, calculations, and decisions behind those outputs?

Those outcomes can be tested without making unsupported assumptions about routing or product architecture.

Three Models Expand the Evaluation Matrix

OpenAI launched GPT-5.6 as a family rather than a single model. Sol is presented as the flagship, Terra as the balanced option for everyday work, and Luna as the option emphasizing speed and affordability.

Model	Stated position in the family	Stated emphasis	What organizations should evaluate
GPT-5.6 Sol	Flagship model	Highest capability	Whether additional capability materially improves difficult, high-value tasks
GPT-5.6 Terra	Balanced model	Everyday capability and efficiency	Whether it delivers consistent results across routine knowledge work
GPT-5.6 Luna	Fastest and most affordable version	Speed and cost	Whether faster or less expensive output remains accurate enough for the intended task

This three-tier structure creates possible deployment and product-design choices, but the supplied information does not establish how Microsoft 365 Copilot will use those choices. Microsoft might expose model selection, make selections within the product, use only part of the family on particular surfaces, or change availability over time. Until Microsoft documents that behavior, none of those possibilities should be presented as current product fact.
The model family nevertheless changes how organizations should think about evaluation. “Use the most capable model” is not always a sufficient policy. A model’s business value depends on the task, the cost of an incorrect result, the time required for review, and the amount of human correction needed before the work can be approved.
A lightweight rewriting task and a consequential financial analysis should not necessarily have the same acceptance criteria. The first may prioritize speed and tone consistency. The second should prioritize traceable calculations, explicit assumptions, reproducibility, and review by a qualified person.
Nikunj Handa, OpenAI’s API Product lead, described the economic goal as helping organizations get “more useful work from every token.” That framing puts the emphasis on completed work rather than raw output volume. Token efficiency is useful only when it preserves the context needed to produce a correct result.
Fewer tokens are not inherently better. A short analysis that omits a critical worksheet, source document, exception, or business constraint can be less valuable than a longer process that includes the necessary evidence. The proper measure is useful, verifiable completion—not minimum computation in isolation.
GPT-5.6 should therefore be judged on more than model-family positioning. OpenAI must demonstrate that the models produce materially better results, while Microsoft 365 Copilot must demonstrate those improvements within the applications and workflows named in the announcement.

Word, Excel, and PowerPoint Are Becoming AI Workspaces

The integration’s most immediate promises concern applications employees already use. In Word, GPT-5.6 is expected to create, edit, and refine documents with fewer prompts. In Excel, the promise is deeper analysis with more efficient token usage. In PowerPoint, the model is expected to generate higher-quality presentations requiring less manual cleanup.
These improvements may initially sound incremental because Copilot already performs document, spreadsheet, and presentation tasks. The proposed difference is a reduction in interaction and correction costs: fewer follow-up prompts, less manual repair, and fewer outputs that appear polished but leave the difficult final work to the user.

Word: measure the distance to an approved document

A useful Word evaluation should go beyond asking the model to draft a generic paragraph. Organizations should test whether it can work from approved source material, follow layered instructions, preserve required terminology, and maintain consistency across a long document.
A realistic test might ask Copilot to turn a collection of approved materials into a structured proposal while preserving product names, contractual language, tone, and audience requirements. Reviewers should then identify:

Unsupported claims introduced by the model.
Approved facts that were omitted or distorted.
Inconsistent terminology across sections.
Changes to text that was supposed to remain untouched.
Citations or source references that cannot be traced.
The number of prompts and manual edits required to reach approval.

The central metric is edit distance: how much human work separates the initial output from a document that the organization is willing to publish, send, sign, or archive.
Fewer prompts are valuable only if they lead to fewer substantive corrections. A one-prompt document containing hidden factual errors is not a productivity improvement over a three-prompt document whose claims remain grounded in approved material.

Excel: make every important assumption auditable

Excel is a harder test because a persuasive narrative can conceal an incorrect range, misunderstood label, missing value, broken formula, or invalid business assumption. Enterprise workbooks are rarely as clean as demonstration data. They often contain historical conventions, renamed categories, manually entered exceptions, hidden sheets, linked sources, and logic known only to their creators.
Deeper analysis should therefore be evaluated through reproducibility. When Copilot identifies a trend or recommends an explanation, an analyst should be able to determine:

Which worksheets, tables, ranges, or fields were used.
Whether filters or exclusions affected the result.
How missing and duplicate values were handled.
Which formulas or transformations produced the conclusion.
Whether the model separated observed correlation from a proposed cause.
Whether another reviewer can reproduce the result independently.

Token efficiency may help if the system focuses on relevant workbook context instead of processing unnecessary information. It becomes a liability if the model silently excludes data that changes the conclusion.
WindowsForum’s practical recommendation is simple: do not accept an Excel narrative because it sounds confident. Audit the ranges, formulas, filters, assumptions, and intermediate calculations. For consequential decisions, require an analyst to reproduce the result before the conclusion leaves the working team.

PowerPoint: judge the argument, not just the appearance

PowerPoint is the most visible test because weak results are immediately apparent. An AI-generated deck can be grammatically clean and visually polished while still being repetitive, unsupported, badly paced, or disconnected from the presenter’s objective.
Higher-quality presentation generation should be measured across at least four dimensions:

Accuracy: Do the claims match approved source material?
Narrative: Does the deck build a coherent argument rather than repeat summary points?
Evidence: Are charts, quotations, and figures traceable to reliable inputs?
Usability: How much redesign and rewriting is needed before presentation?

Reviewers should also compare the deck with the source package. A presentation that looks better than the underlying evidence warrants more scrutiny, not less. Visual polish can make unsupported conclusions feel authoritative.
Nitin Agrawal, Microsoft Copilot & Agents Core President, said customers will be able to produce more polished outputs across Word, Excel, PowerPoint, Cowork, and Copilot Chat. “More polished” is a practical claim that organizations can test. It should mean that less human effort is required to transform an AI-generated artifact into approved work—not merely that the first output contains more text or decoration.

Cowork Makes the Upgrade Operationally Significant

GPT-5.6’s largest consequences may emerge in Copilot Chat and Cowork, where the stated promise is support for more capable, complex, multi-step workflows.
The supplied information does not establish Cowork’s detailed product design. It does not confirm how plans are displayed, whether tasks continue in the background, which tools can be used, how approvals are requested, or how administrators configure workflow boundaries. Those details require Microsoft product documentation before they can be described as current behavior.
The broader significance of multi-step work is still clear. A single generated paragraph has a limited failure radius. A workflow that uses one output as the input for another can propagate an early error across several artifacts.
Consider a quarterly business review as an evaluation scenario. An AI workflow might be asked to help analyze performance data, summarize source documents, prepare a written briefing, and create a presentation. If the initial analysis misclassifies a product line, later documents may repeat that mistake as though it were an established fact.
Each additional step increases the need for checkpoints. Practical evaluation should ask:

Can reviewers identify the source material used at each stage?
Are calculations and transformations visible enough to audit?
Does the workflow stop when required information is missing?
Does it expose conflicts between sources instead of silently choosing one?
Can a reviewer reject or revise an intermediate result?
Is approval required before a consequential action leaves the working environment?

These are evaluation criteria, not claims about Cowork’s current controls. Microsoft has not provided enough information in the supplied fact set to confirm how the product handles these requirements.
The distinction between preparation and execution is especially important. Drafting an external message is different from sending it. Proposing a spreadsheet change is different from applying it to an authoritative record. Preparing a decision memo is different from authorizing the decision.
Organizations assessing multi-step workflows should preserve those distinctions even when the AI result appears accurate. Convenience should not erase approval boundaries.

Ultra Mode Treats Intelligence as a Small Team

OpenAI also introduced ultra mode, which coordinates four agents simultaneously for complex tasks. Instead of relying on one model instance to work through the entire assignment, the approach distributes work among several agents and combines their contributions.
The useful analogy is a temporary project team. Different agents could investigate separate aspects of a problem, test assumptions, compare evidence, or contribute to a combined result. The precise division depends on implementation, but the central premise is that some assignments benefit from parallel work.
Multiple agents do not automatically create independent verification. They can share the same missing context, repeat one another’s assumptions, duplicate effort, or produce conflicting conclusions that are combined too confidently.
For workplace evaluation, the important issue is therefore not the number of agents. It is whether users can understand the resulting work well enough to review it. A credible multi-agent result should make it possible to determine:

How the assignment was divided.
Which materials informed each part.
Where assumptions entered the process.
Whether agents reached conflicting conclusions.
How conflicts were resolved.
Which evidence supports the final output.

Without that visibility, adding agents can increase cost and complexity without increasing trust.
Ultra mode nevertheless illustrates the direction of high-end AI work. The competitive frontier is moving beyond a smarter conversation toward systems that can coordinate multiple reasoning processes around a shared objective. Microsoft 365 customers should evaluate that direction through evidence quality, reproducibility, and approval discipline rather than through agent counts alone.

GPT-5.6 Arrives With Claims That Need Workplace Evidence

OpenAI describes GPT-5.6 Sol as its strongest model for programming and cybersecurity and reports improvements across scientific research, data analysis, and document workflows. Those claims help explain the flagship model’s positioning, but they do not settle whether GPT-5.6 will improve a particular Microsoft 365 deployment.
Highly specific benchmark comparisons require an on-topic primary source and enough methodological detail to interpret the result. The available information is not sufficient to publish the claim that GPT-5.6 Sol defeated a named competitor on Agents’ Last Exam, so that comparison should not be used as evidence here.
Benchmarks and workplace results answer different questions. A benchmark generally provides defined inputs, expected outcomes, and a scoring method. Office work includes incomplete records, contradictory files, unclear requests, undocumented conventions, and stakeholder preferences that cannot be reduced to a single correct answer.
The Microsoft 365 integration may therefore be more revealing than an isolated model score. A user might ask Excel to explain a revenue decline without mentioning that a product category changed names during the reporting period. A model can perform correct arithmetic on the wrong grouping and produce a persuasive but invalid business interpretation.
The strongest evidence will come from repeatable workplace outcomes:

Can Word preserve factual consistency through several rounds of revision?
Can Excel expose calculations and assumptions for independent review?
Can PowerPoint trace important claims to approved evidence?
Can Copilot Chat and Cowork identify missing inputs before continuing?
Do results remain reliable across different users and imperfect files?
How often must a knowledgeable reviewer correct a substantive mistake?

Organizations should avoid evaluating GPT-5.6 only through isolated “wow” prompts. A strong first draft is encouraging, but enterprise value depends on repeatability. The relevant question is not whether the system can produce one impressive result. It is whether it can deliver dependable results across ordinary, messy, and adversarial cases.

Safety Claims Do Not Replace Deployment Review

OpenAI says GPT-5.6 includes its most reliable safety system to date, with model safeguards, monitoring, prompt screening, and additional restrictions for high-risk use cases.
The model also underwent review by U.S. authorities, according to the supplied information. That fact should be interpreted narrowly. It does not establish government certification for Microsoft 365 deployment, approval for regulated workloads, universal model safety, or security for any individual tenant.
A provider’s model-level safety measures and an organization’s deployment controls address different risks. Model safeguards may reduce certain harmful outputs. They do not determine whether an employee has access to an internal document, whether a spreadsheet contains sensitive data, whether a source is approved, or whether a user is authorized to act on a generated recommendation.
Organizations should consequently test GPT-5.6 within their own information and decision environment. Important questions include whether the model reveals sensitive information to an inappropriate audience, relies on outdated source files, creates unsupported claims, or encourages action without adequate review.
The level of review should track the consequence of the task. Low-risk drafting may need ordinary editorial review. Work involving external communication, financial reporting, personnel matters, security operations, legal commitments, regulated data, or destructive changes should face stricter controls and explicit human approval.
These are recommended governance practices, not evidence of specific Microsoft 365 Copilot features. The supplied facts do not confirm tenant logging capabilities, approval controls, permission models, or administrative configuration options for GPT-5.6.
The central safety question is not merely whether the model refuses prohibited text. It is whether an organization can detect a plausible but incorrect result before that result influences people, records, systems, or external communications.

ChatGPT Work Extends the Thesis Beyond Office

OpenAI’s GPT-5.6 rollout also includes ChatGPT Work, an agent intended for complex, multi-step tasks. The supplied information says it can work with user files and integrate with enterprise services including Google Drive, Slack, Microsoft Teams, and Salesforce.
That creates an increasingly complex enterprise market. GPT-5.6 strengthens Microsoft 365 Copilot while OpenAI also applies the model family to its own workplace product. Customers may therefore encounter overlapping tools that can participate in document creation, analysis, communication, and workflow execution.
The available facts do not establish ChatGPT Work’s tenant controls, licensing, regional availability, administrative model, data boundaries, or exact connector behavior. Those issues require product-specific documentation and should not be inferred from the existence of an integration.
For organizations, the strategic question is broader than which model is preferred. It is which product is authorized to use which data for which purpose. Before adopting overlapping workplace agents, decision-makers should define:

Which repositories each product may access.
Which data classifications are permitted.
Which outputs may become organizational records.
Who reviews consequential work.
Who owns and maintains recurring workflows.
How access is removed when an employee changes roles.
What evidence is required before a generated conclusion is acted upon.

The workplace agent that produces the most impressive demonstration may not be the one that fits an organization’s requirements. Long-term value will depend on whether the product can be evaluated, governed, and incorporated into accountable business processes.

Recommended Governance Checklist Before Broad Deployment

The GPT-5.6 announcement should be treated as a reason to prepare an evaluation plan, not as proof that administrators must immediately change a Microsoft 365 setting.
The supplied information does not provide a verified Microsoft 365 Admin Center path, required role, license SKU, configuration name, regional rollout schedule, tenant control, or enablement procedure. It also does not establish Microsoft’s official guidance for administering GPT-5.6.
Administrators should consult current Microsoft documentation and tenant-specific communications before attempting configuration changes. Until those materials are available, the following checklist represents recommended practice for evaluating a more capable AI service—not current Microsoft product instructions.

Recommended evaluation and governance checklist

Define the intended use cases. Identify the Word, Excel, PowerPoint, Copilot Chat, and Cowork tasks the organization expects GPT-5.6 to improve. Avoid broad goals such as “increase productivity” without measurable outcomes.
Separate low-risk assistance from consequential work. Classify drafting, summarization, analysis, external communication, record changes, financial decisions, security operations, and other actions according to their potential impact.
Test with representative material. Use realistic document structures, workbook complexity, templates, terminology, and workflow constraints. Sanitized test data should preserve the difficulty of the real task.
Compare outputs with approved sources. Require reviewers to verify important statements against authoritative documents rather than accepting a polished answer at face value.
Audit spreadsheet assumptions. Check ranges, formulas, filters, units, missing values, exclusions, transformations, and proposed explanations. Require independent reproduction for consequential analysis.
Measure edit distance. Record how many prompts, corrections, factual repairs, formatting changes, and expert interventions are needed before an output can be approved.
Test failure behavior. Remove a source, introduce conflicting figures, rename a worksheet, include outdated material, or provide ambiguous instructions. Determine whether the system identifies the problem or continues confidently.
Require approval before consequential actions. Preserve a human decision point before external messages, financial commitments, personnel decisions, security changes, destructive edits, or updates to authoritative records.
Document unresolved product questions. Track confirmed information about availability, licensing, regional rollout, model selection, administrative controls, and enablement steps as Microsoft publishes it. Do not substitute assumptions for documentation.
Establish ownership. Assign responsibility for test design, source approval, output review, incident handling, workflow maintenance, and the decision to expand or suspend use.
Measure completion quality and correction cost. Compare GPT-5.6-assisted work with the current process, including time saved, errors introduced, expert review required, tasks completed successfully, and failures that could have produced business harm.

GPT-5.6’s preferred status in Microsoft 365 Copilot is significant, but the announcement is the beginning of the deployment question rather than its answer. The verified promises—fewer prompts in Word, deeper and more token-efficient Excel analysis, higher-quality PowerPoint output, and more capable Copilot Chat and Cowork workflows—are specific enough to test.
Organizations should run those tests against approved source material, audit spreadsheet assumptions, measure the path from generated output to accepted work, and require human approval before consequential actions. If GPT-5.6 consistently reduces that path without weakening accuracy or accountability, the preferred model designation will translate into practical value. If it merely produces more polished output that still demands extensive correction, the most important improvement will remain the discipline of the people evaluating it.

References

Primary source: incrypted
Published: 2026-07-11T08:42:07.457135

OpenAI Integrated GPT-5.6 into Microsoft 365 Copilot

Sam Altman has announced the integration of the flagship GPT-5.6 model into Microsoft 365 Copilot. What can the new ultra mode do, how will the three-model lineup supercharge Word and Excel, and why will this change office work forever? Details on the technological breakthrough.

incrypted.com
Official source: help.openai.com

GPT-5.6 in ChatGPT | OpenAI Help Center

Learn how GPT-5.6 Sol works in ChatGPT, which options are available by plan, and how usage limits and availability work.

help.openai.com
Official source: openai.com

Previewing GPT-5.6 Sol: a next-generation model | OpenAI

OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.

openai.com
Official source: learn.microsoft.com

OpenAI as a subprocessor in Microsoft Online Services | Microsoft Learn

Learn about using OpenAI as a subprocessor in Microsoft Online Services, including how to enable and disable in the Microsoft 365 admin center.

learn.microsoft.com
Related coverage: techcrunch.com

OpenAI says GPT 5.6 is the 'preferred model' for Microsoft Copilot 365 amid breakup chatter | TechCrunch

OpenAI's new family of models will continue to power Microsoft's suite of workplace and productivity apps.

techcrunch.com
Related coverage: tech.yahoo.com

OpenAI GPT-5.6 is now the preferred model for Microsoft 365 Copilot

The model, which only received broad U.S. regulatory clearance this week, is now rolling out across Word, Excel, PowerPoint, Chat, and Cowork

tech.yahoo.com

Official source: deploymentsafety.openai.com

GPT-5.6 System Card - OpenAI Deployment Safety Hub

GPT-5.6 is a new family of three models: Sol, our new flagship model; Terra, a capable lower-cost option; and Luna, our fastest and most cost-efficient model. The safeguards we have built for this launch—our most robust yet—are built to deliver these models safely and at scale, around the world.

deploymentsafety.openai.com
Related coverage: ddazcdn01.z8.web.core.windows.net

Copilot for Microsoft365 Partner FAQ

PDF document

ddazcdn01.z8.web.core.windows.net
Official source: cdn.openai.com

genebench pro

PDF document

cdn.openai.com
Official source: cdn-dynmedia-1.microsoft.com

Microsoft Copilot for Microsoft 365

Microsoft Copilot for Microsoft 365

cdn-dynmedia-1.microsoft.com
Official source: marketingassets.microsoft.com

Microsoft 365 Copilot Experience Guide

PDF document

marketingassets.microsoft.com
Related coverage: covenant.global

1186904263

High angle shot of a group of businesspeople talking to each other while standing in an office

covenant.global
Related coverage: tomsguide.com

The Trump administration just quietly changed how OpenAI will launch its next model | Tom's Guide

The Trump administration reportedly asked OpenAI to limit GPT-5.6's launch to government-approved partners. It could mark the beginning of a new era for frontier AI.

www.tomsguide.com
Related coverage: tomshardware.com

OpenAI's ChatGPT-5.6 gets the same banhammer treatment as Anthropic’s Mythos from the federal government — source says that Washington cautioned OpenAI against releasing the model without receiving approval | Tom's Hardware

The U.S. government wants to ensure that its latest, most advanced AI tools can't be used against it.

www.tomshardware.com
Related coverage: windowscentral.com

Microsoft Drops GPT-5 Into Copilot—No Subscription Needed | Windows Central

OpenAI's much-hyped GPT-5 AI model is now available as part of Microsoft Copilot. Here's what you need to know.

www.windowscentral.com
Official source: microsoft.com

Copilot's agentic capabilities in Word, Excel, and PowerPoint are generally available | Microsoft 365 Blog

Copilot's newest features are now generally available as the default experience across all Microsoft 365 subscriptions.

www.microsoft.com
Official source: blogs.microsoft.com

Introducing Microsoft 365 Copilot – your copilot for work - The Official Microsoft Blog

Humans are hard-wired to dream, to create, to innovate. Each of us seeks to do work that gives us purpose — to write a great novel, to make a discovery, to build strong communities, to care for the sick. The urge to connect to the core of our work lives in all of us. But...

blogs.microsoft.com
Official source: techcommunity.microsoft.com

https://techcommunity.microsoft.com/blog/microsoft365copilotblog/introducing-word-excel-and-powerpoint-agents-in-microsoft-365-copilot/4470604/replies/4505512
Official source: news.microsoft.com

Satya Nadella The Future of Work with AI

PDF document

news.microsoft.com

Microsoft 365 tool	Benefit stated in the supplied report	Practical enterprise test
Word	Faster drafting and editing with fewer prompts	Preserving facts, tone, structure, and source meaning through revisions
Excel	More efficient data analysis and quicker insights	Distinguishing supported findings from misleading or incomplete interpretations
PowerPoint	More visually refined presentations with minimal manual input	Converting source material into a coherent narrative without inventing support
Cowork	Streamlined collaboration and task completion	Completing a defined task accurately while preserving instructions and review boundaries

Navigation section

GPT-5.6 Becomes Microsoft 365 Copilot’s Preferred Model July 9

Microsoft Is Upgrading the Work, Not Just the Chatbot​

The Enterprise Test Is the Cost of Reaching an Acceptable Result​

Four Applications, Four Different Tests​

Word​

Excel​

PowerPoint​

Cowork​

Output per Token Is Not the Same as Customer Savings​

The Model Is One Part of the Copilot Experience​

The Restricted Preview Establishes a Narrow Cybersecurity Fact Pattern​

Timeline​

A Preferred Model Is Not Proof of Universal Availability​

One Strong Verification Rule: Review According to Impact​

A Bounded Admin Checklist for Measuring GPT-5.6​

1. Identify the test users​

2. Choose fixed application tasks​

3. Retain the source files and prompts​

4. Define review criteria before testing​

5. Record the work required​

6. Compare before and after availability​

7. Report results by application and task​

The Upgrade Gives IT a Chance to Measure Copilot Honestly​

Update: OpenAI frames GPT-5.6 Copilot role as partnership continuity (July 10, 2026)​

Update: GPT-5.6 family adds Terra and Luna variants (July 10, 2026)​

Update: PowerPoint rollout adds branded-template improvements and model comparison (July 10, 2026)​

Update: GPT-5.6 adds automatic routing and manual model selection (July 10, 2026)​

References​

AI

OpenAI Reassures the Market Without Explaining the Router​

The Breakup Story Was Always Too Simple​

Microsoft Has Turned Copilot Into a Cost-Allocation Problem​

“Preferred” Is a Product Claim, Not an Architecture Diagram​

The Office Franchise Is Where AI Economics Get Real​

The Market Heard Reassurance; IT Should Hear Complexity​

Timeline​

This Is the Multi-Model Future Arriving Through the Side Door​

Copilot Customers Need Model Governance, Not Model Gossip​

Action checklist for admins​

The Real Contest Is Over Who Owns the Work Interface​

The “Preferred Model” Label Buys Time, Not Finality​

What WindowsForum Readers Should Watch Next​

References​

AI

AI

AI

GPT‑5.6 in Microsoft 365 Copilot: What the Model Change Means for Users and IT Admins​

Microsoft Is Announcing Copilot Integration, Not a ChatGPT Window​

What Users Can Do Now​

General prompt examples​

The Upgrade Matters Where Answers Become Files​

Word, PowerPoint, and Cowork Require Focused Review​

Excel Is Where Polish Meets the Audit Trail​

What Admins Should Verify​

“More Polished” Raises the Standard for Human Review​

What This Upgrade Changes—and What It Does Not​

Deployment timeline based on the supplied information​

References​

AI

What Changed: GPT-5.6 Is Preferred, Not Guaranteed for Every Request​

Microsoft 365 Copilot Is a Multi-Model Product​

Analysis: What multi-model Copilot may imply​

Sol, Terra, and Luna Divide GPT-5.6 Into Three Positions​

Analysis: Why the three positions matter​

Word, Excel, and PowerPoint Need Measurable Reductions in Rework​

Cowork Raises the Standard from Drafting to Completion​

Security Claims Must Remain Separate from Output Claims​

What Admins Should Verify in Their Pilot Environment​

1. Verify selector or automatic-option presence​

2. Verify enabled Copilot surfaces​

3. Verify access scope​

4. Establish a controlled test corpus​

5. Assign approval and review owners​

Compact Pilot Benchmark Template​

Microsoft and OpenAI Are Cooperating While Microsoft Builds Alternatives​

Analysis: What the relationship means for customers​

The Missing Traffic Numbers Matter More Than the Launch Label​

IT Should Benchmark Workflows, Not Model Personalities​

Recommended decision metrics​

Microsoft Is Upgrading the Work, Not Just the Chatbot

The Enterprise Test Is the Cost of Reaching an Acceptable Result

Four Applications, Four Different Tests

Word

Excel

PowerPoint

Cowork

Output per Token Is Not the Same as Customer Savings

The Model Is One Part of the Copilot Experience

The Restricted Preview Establishes a Narrow Cybersecurity Fact Pattern

Timeline

A Preferred Model Is Not Proof of Universal Availability

One Strong Verification Rule: Review According to Impact

A Bounded Admin Checklist for Measuring GPT-5.6

1. Identify the test users

2. Choose fixed application tasks

3. Retain the source files and prompts

4. Define review criteria before testing

5. Record the work required

6. Compare before and after availability

7. Report results by application and task

The Upgrade Gives IT a Chance to Measure Copilot Honestly

Update: OpenAI frames GPT-5.6 Copilot role as partnership continuity (July 10, 2026)

Update: GPT-5.6 family adds Terra and Luna variants (July 10, 2026)

Update: PowerPoint rollout adds branded-template improvements and model comparison (July 10, 2026)

Update: GPT-5.6 adds automatic routing and manual model selection (July 10, 2026)

References

OpenAI Reassures the Market Without Explaining the Router

The Breakup Story Was Always Too Simple

Microsoft Has Turned Copilot Into a Cost-Allocation Problem

“Preferred” Is a Product Claim, Not an Architecture Diagram

The Office Franchise Is Where AI Economics Get Real

The Market Heard Reassurance; IT Should Hear Complexity

Timeline

This Is the Multi-Model Future Arriving Through the Side Door

Copilot Customers Need Model Governance, Not Model Gossip

Action checklist for admins

The Real Contest Is Over Who Owns the Work Interface

The “Preferred Model” Label Buys Time, Not Finality

What WindowsForum Readers Should Watch Next

References

GPT‑5.6 in Microsoft 365 Copilot: What the Model Change Means for Users and IT Admins

Microsoft Is Announcing Copilot Integration, Not a ChatGPT Window

What Users Can Do Now

General prompt examples

The Upgrade Matters Where Answers Become Files

Word, PowerPoint, and Cowork Require Focused Review

Excel Is Where Polish Meets the Audit Trail

What Admins Should Verify

“More Polished” Raises the Standard for Human Review

What This Upgrade Changes—and What It Does Not

Deployment timeline based on the supplied information

References

What Changed: GPT-5.6 Is Preferred, Not Guaranteed for Every Request

Microsoft 365 Copilot Is a Multi-Model Product

Analysis: What multi-model Copilot may imply

Sol, Terra, and Luna Divide GPT-5.6 Into Three Positions

Analysis: Why the three positions matter

Word, Excel, and PowerPoint Need Measurable Reductions in Rework

Cowork Raises the Standard from Drafting to Completion

Security Claims Must Remain Separate from Output Claims

What Admins Should Verify in Their Pilot Environment

1. Verify selector or automatic-option presence

2. Verify enabled Copilot surfaces

3. Verify access scope

4. Establish a controlled test corpus

5. Assign approval and review owners

Compact Pilot Benchmark Template

Microsoft and OpenAI Are Cooperating While Microsoft Builds Alternatives

Analysis: What the relationship means for customers

The Missing Traffic Numbers Matter More Than the Launch Label

IT Should Benchmark Workflows, Not Model Personalities

Recommended decision metrics

A Practical Pilot Timeline

Phase 1: Inventory and baseline

Phase 2: Controlled GPT-5.6 testing

Phase 3: Error analysis

Phase 4: Workflow decision

The Decision Rule: Prefer the Option, Approve the Workflow

References