Microsoft is reportedly moving some Copilot prompts in Excel, Outlook, and other productivity software away from OpenAI and Anthropic models and toward its own MAI model family as of July 2026, according to Bloomberg reporting summarized by SiliconANGLE and other outlets. The shift is not a divorce from OpenAI, nor even a wholesale model replacement. It is something more revealing: Microsoft is beginning to treat frontier AI as a margin problem, not just a magic trick. For Windows users and IT departments, that may matter more than another benchmark win.
The great irony of Microsoft’s generative AI era is that the company which sold Wall Street on the dream of AI everywhere now appears to be doing what every Excel user eventually does: auditing the bill.
Bloomberg reported this week that Microsoft has started routing “tens of thousands” of prompts in products such as Excel and Outlook through its own MAI models instead of the highest-end systems from OpenAI and Anthropic. SiliconANGLE’s write-up frames the move bluntly: Microsoft is reportedly leaning on its in-house models to cut costs, even while acknowledging that they are not necessarily as sophisticated as the leading frontier systems.
That distinction matters. Microsoft is not saying that MAI is suddenly better than OpenAI’s or Anthropic’s best models across the board. It is saying, through its product behavior, that many enterprise AI requests may not need the best model in the world. They need a model that is good enough, fast enough, cheap enough, and controllable enough to run at Microsoft 365 scale.
That is a very different kind of AI competition from the leaderboard wars of 2023 and 2024. It is less glamorous, less tweetable, and probably more important.
That model never made much economic sense at enterprise scale. Microsoft 365 is not a boutique chatbot. It is a high-volume software estate where even small inference costs can turn ugly when multiplied across Outlook, Excel, Word, Teams, PowerPoint, Windows, GitHub, and Azure customers.
This is why the Bloomberg report lands as more than vendor gossip. If Microsoft is moving even a small fraction of Copilot workloads onto MAI, it is admitting that the economic architecture of AI software must be layered. Some requests merit a heavyweight model. Many do not.
A user asking Outlook to shorten an email probably does not need the same model Microsoft would use to debug a distributed systems failure. An Excel user asking for a formula suggestion may benefit more from predictable latency and low cost than from a system with world-class abstract reasoning. The future of AI productivity software is likely not one model to rule them all, but a routing layer that decides how expensive your question is allowed to become.
Those details are not just spec-sheet decoration. A sparse mixture-of-experts model that activates fewer parameters per request is designed to reduce inference cost while preserving enough capability for demanding tasks. In plain English, Microsoft wants a model that can do serious work without behaving like a private jet meter is running every time someone asks Copilot to summarize a meeting.
The company has positioned MAI-Thinking 1 as competitive in certain coding evaluations, including comparisons with Anthropic’s Claude Opus 4.6 in blind testing cited by Microsoft-focused coverage. But the more important point is not whether MAI wins a benchmark on Tuesday. It is whether Microsoft can reliably classify workloads so that MAI handles the routine cases while OpenAI, Anthropic, or other frontier systems remain available for jobs that genuinely need them.
That is the shape of a mature AI business. The expensive model becomes the escalation path, not the default plumbing.
But dependence is different from partnership. Microsoft’s deal with OpenAI reportedly runs until 2032, and even if the economics are favorable compared with public API pricing, the costs still accumulate when Copilot is asked to become a default interface across the company’s software universe. A discounted expensive thing can still be expensive.
There is also a strategic control issue. If Microsoft wants to make AI a native layer of Windows and Microsoft 365, it cannot afford to have every critical product decision mediated by another lab’s roadmap, pricing, safety policies, rate limits, or model availability. Owning more of the model stack gives Microsoft leverage.
That leverage may be useful even if Microsoft continues using OpenAI heavily. The existence of MAI changes the negotiating posture. OpenAI is no longer the only answer to every internal product team asking, “What model do we use?”
That is unusually plain language for a senior executive discussing an AI supplier. It also reflects a broader market mood. The enterprise AI land rush has entered the invoice stage, and executives who once asked why their companies were not using AI everywhere are now asking why AI everywhere costs so much.
Anthropic’s Claude models have built a strong reputation with developers and enterprise users, particularly for coding, writing, long-context work, and cautious behavior. That reputation gives Anthropic pricing power. But pricing power can become a target when hyperscalers decide they need cheaper alternatives at massive volume.
Microsoft’s reported shift does not mean Anthropic is doomed inside enterprise software. It does mean that a premium model provider must keep proving that the premium is justified. In many tasks, “best” is a luxury adjective. In enterprise procurement, “good enough at one-tenth the cost” can be a strategy.
The first test gets the demos. The second test decides the business.
A Copilot feature that delights users but costs too much to run is not a product; it is a subsidy. Microsoft can tolerate subsidies during a land grab, especially when the prize is reshaping Office, Windows, GitHub, and Azure. But subsidies become harder to defend once customers start asking for proof that AI is worth the seat price and investors start asking when AI gross margins resemble software gross margins.
That is why model routing is likely to become one of the most important invisible systems in Microsoft’s AI stack. The user sees a Copilot button. Behind it, Microsoft sees a cost tree: prompt length, task type, latency target, tenant policy, data sensitivity, model availability, expected output quality, and escalation logic.
The winning AI assistant may not be the one with the single smartest model. It may be the one with the best dispatcher.
But model substitution can still show up in the experience. Users may notice differences in tone, formatting, refusal behavior, reasoning depth, hallucination patterns, or how well Copilot handles edge cases. IT administrators may notice more predictable pricing and performance before end users notice any model change at all.
For Windows enthusiasts, this also hints at where local and cloud AI may be headed. Microsoft has already been pushing AI PCs, NPUs, and on-device workloads, but the reported MAI shift is mostly about cloud inference economics. The underlying principle is the same: not every AI task deserves the biggest remote model.
A future Windows system may use a local small model for simple actions, a Microsoft-hosted MAI model for productivity tasks, and a premium frontier model for complex reasoning. That would make Copilot less like a chatbot and more like an operating system service. The user would not ask which engine is running. The system would choose.
On the other hand, model routing creates a new transparency problem. If the answer to “What model processed this data?” changes by workload, tenant, region, product, and cost-optimization rule, administrators will need better documentation than marketing pages can provide. In regulated industries, the model used is not a trivia detail. It can affect auditability, data handling, reproducibility, and risk assessments.
Microsoft will need to explain not just whether customer data is protected, but how model selection works under enterprise controls. Can a tenant force specific models? Can it block third-party models entirely? Can it require Microsoft-owned models for certain workloads? Can it produce logs showing which model processed a prompt?
These questions will become more urgent as Copilot moves deeper into Office documents, email, Teams meetings, Power Platform workflows, endpoint management, and security products. AI governance is not only about whether a model says something dangerous. It is also about whether an organization knows what system touched its data in the first place.
The early phase of generative AI rewarded brute force. Use the strongest model. Add more context. Generate more drafts. Turn on the agent. Let the model think longer. If the output was impressive, the cost could be rationalized as innovation.
That phase is ending. The next phase is about cost-aware intelligence. Enterprises want smaller models for narrow tasks, cheaper models for high-volume tasks, and frontier models only where the incremental capability changes the outcome.
This is not a retreat from AI. It is the industrialization of AI. Every major computing platform eventually learns the same lesson: performance matters, but performance per dollar matters more.
Building MAI gives Microsoft a way to say, internally and externally, that it is not merely the cloud host for someone else’s brain. That matters to developers, customers, regulators, and Wall Street. It also matters to Microsoft’s own product groups, which need confidence that the company can ship AI features on its own schedule.
The reported use of MAI in Excel and Outlook is symbolically important because those are not side projects. They are core Microsoft products. If Microsoft is comfortable using its own models there, even for a small share of traffic, it is making a statement about trust.
The statement is not “we no longer need OpenAI.” It is “we need options.” In platform strategy, options are power.
Microsoft’s AI Strategy Has Reached the Spreadsheet Phase
The great irony of Microsoft’s generative AI era is that the company which sold Wall Street on the dream of AI everywhere now appears to be doing what every Excel user eventually does: auditing the bill.Bloomberg reported this week that Microsoft has started routing “tens of thousands” of prompts in products such as Excel and Outlook through its own MAI models instead of the highest-end systems from OpenAI and Anthropic. SiliconANGLE’s write-up frames the move bluntly: Microsoft is reportedly leaning on its in-house models to cut costs, even while acknowledging that they are not necessarily as sophisticated as the leading frontier systems.
That distinction matters. Microsoft is not saying that MAI is suddenly better than OpenAI’s or Anthropic’s best models across the board. It is saying, through its product behavior, that many enterprise AI requests may not need the best model in the world. They need a model that is good enough, fast enough, cheap enough, and controllable enough to run at Microsoft 365 scale.
That is a very different kind of AI competition from the leaderboard wars of 2023 and 2024. It is less glamorous, less tweetable, and probably more important.
The Frontier Model Was Never Going to Handle Every Calendar Invite
The consumer narrative around AI has long implied that every prompt deserves the most powerful model available. Ask for a summary, a chart, a rewrite, or a scheduling suggestion, and somewhere behind the curtain sits the same sort of premium reasoning engine that might also write code, solve math problems, or analyze a contract.That model never made much economic sense at enterprise scale. Microsoft 365 is not a boutique chatbot. It is a high-volume software estate where even small inference costs can turn ugly when multiplied across Outlook, Excel, Word, Teams, PowerPoint, Windows, GitHub, and Azure customers.
This is why the Bloomberg report lands as more than vendor gossip. If Microsoft is moving even a small fraction of Copilot workloads onto MAI, it is admitting that the economic architecture of AI software must be layered. Some requests merit a heavyweight model. Many do not.
A user asking Outlook to shorten an email probably does not need the same model Microsoft would use to debug a distributed systems failure. An Excel user asking for a formula suggestion may benefit more from predictable latency and low cost than from a system with world-class abstract reasoning. The future of AI productivity software is likely not one model to rule them all, but a routing layer that decides how expensive your question is allowed to become.
MAI Is Microsoft’s Bid to Own the Middle of the Model Stack
Microsoft’s MAI family gives the company a way to occupy the middle ground between cheap commodity models and frontier systems from OpenAI, Anthropic, Google, and others. The company recently introduced several MAI models, including MAI-Thinking 1, which Microsoft describes as a midsized reasoning model with 35 billion active parameters and a 256,000-token context window.Those details are not just spec-sheet decoration. A sparse mixture-of-experts model that activates fewer parameters per request is designed to reduce inference cost while preserving enough capability for demanding tasks. In plain English, Microsoft wants a model that can do serious work without behaving like a private jet meter is running every time someone asks Copilot to summarize a meeting.
The company has positioned MAI-Thinking 1 as competitive in certain coding evaluations, including comparisons with Anthropic’s Claude Opus 4.6 in blind testing cited by Microsoft-focused coverage. But the more important point is not whether MAI wins a benchmark on Tuesday. It is whether Microsoft can reliably classify workloads so that MAI handles the routine cases while OpenAI, Anthropic, or other frontier systems remain available for jobs that genuinely need them.
That is the shape of a mature AI business. The expensive model becomes the escalation path, not the default plumbing.
OpenAI Is Still a Partner, but Dependency Is Now a Liability
It would be easy to overstate this story as Microsoft “ditching” OpenAI. That is not what the reported facts support. Microsoft remains deeply tied to OpenAI through Azure, product integration, infrastructure, and years of strategic investment. Its OpenAI partnership remains one of the defining bets of Satya Nadella’s tenure.But dependence is different from partnership. Microsoft’s deal with OpenAI reportedly runs until 2032, and even if the economics are favorable compared with public API pricing, the costs still accumulate when Copilot is asked to become a default interface across the company’s software universe. A discounted expensive thing can still be expensive.
There is also a strategic control issue. If Microsoft wants to make AI a native layer of Windows and Microsoft 365, it cannot afford to have every critical product decision mediated by another lab’s roadmap, pricing, safety policies, rate limits, or model availability. Owning more of the model stack gives Microsoft leverage.
That leverage may be useful even if Microsoft continues using OpenAI heavily. The existence of MAI changes the negotiating posture. OpenAI is no longer the only answer to every internal product team asking, “What model do we use?”
Anthropic’s Price Tag Became the Quiet Villain
The sharper edge of the reporting concerns Anthropic. Microsoft AI chief Mustafa Suleyman reportedly told Bloomberg that Anthropic is “extremely expensive” and that many customers are urgently looking for alternatives. He also said Microsoft pays a lot of money to Anthropic and wants to reduce, and ultimately eliminate, that cost.That is unusually plain language for a senior executive discussing an AI supplier. It also reflects a broader market mood. The enterprise AI land rush has entered the invoice stage, and executives who once asked why their companies were not using AI everywhere are now asking why AI everywhere costs so much.
Anthropic’s Claude models have built a strong reputation with developers and enterprise users, particularly for coding, writing, long-context work, and cautious behavior. That reputation gives Anthropic pricing power. But pricing power can become a target when hyperscalers decide they need cheaper alternatives at massive volume.
Microsoft’s reported shift does not mean Anthropic is doomed inside enterprise software. It does mean that a premium model provider must keep proving that the premium is justified. In many tasks, “best” is a luxury adjective. In enterprise procurement, “good enough at one-tenth the cost” can be a strategy.
Copilot’s Real Test Is Not Intelligence, but Unit Economics
Microsoft Copilot has always faced two tests at once. The first is whether the software is useful enough that users change their habits. The second is whether Microsoft can deliver that usefulness without destroying margins.The first test gets the demos. The second test decides the business.
A Copilot feature that delights users but costs too much to run is not a product; it is a subsidy. Microsoft can tolerate subsidies during a land grab, especially when the prize is reshaping Office, Windows, GitHub, and Azure. But subsidies become harder to defend once customers start asking for proof that AI is worth the seat price and investors start asking when AI gross margins resemble software gross margins.
That is why model routing is likely to become one of the most important invisible systems in Microsoft’s AI stack. The user sees a Copilot button. Behind it, Microsoft sees a cost tree: prompt length, task type, latency target, tenant policy, data sensitivity, model availability, expected output quality, and escalation logic.
The winning AI assistant may not be the one with the single smartest model. It may be the one with the best dispatcher.
Windows and Microsoft 365 Users Will Feel the Shift Indirectly
Most users will not know which model answered a prompt, and Microsoft probably prefers it that way. The Copilot brand is meant to flatten complexity. Users are supposed to ask a question and get an answer, not choose between MAI, GPT, Claude, and a dozen smaller models like they are picking a printer driver in 2006.But model substitution can still show up in the experience. Users may notice differences in tone, formatting, refusal behavior, reasoning depth, hallucination patterns, or how well Copilot handles edge cases. IT administrators may notice more predictable pricing and performance before end users notice any model change at all.
For Windows enthusiasts, this also hints at where local and cloud AI may be headed. Microsoft has already been pushing AI PCs, NPUs, and on-device workloads, but the reported MAI shift is mostly about cloud inference economics. The underlying principle is the same: not every AI task deserves the biggest remote model.
A future Windows system may use a local small model for simple actions, a Microsoft-hosted MAI model for productivity tasks, and a premium frontier model for complex reasoning. That would make Copilot less like a chatbot and more like an operating system service. The user would not ask which engine is running. The system would choose.
Enterprise IT Gets a New Governance Problem
For sysadmins and enterprise architects, Microsoft’s move cuts both ways. On one hand, more in-house Microsoft models could simplify procurement and reduce exposure to third-party model providers. If Copilot workloads stay inside Microsoft’s cloud and are handled by Microsoft-controlled models, some compliance conversations may become easier.On the other hand, model routing creates a new transparency problem. If the answer to “What model processed this data?” changes by workload, tenant, region, product, and cost-optimization rule, administrators will need better documentation than marketing pages can provide. In regulated industries, the model used is not a trivia detail. It can affect auditability, data handling, reproducibility, and risk assessments.
Microsoft will need to explain not just whether customer data is protected, but how model selection works under enterprise controls. Can a tenant force specific models? Can it block third-party models entirely? Can it require Microsoft-owned models for certain workloads? Can it produce logs showing which model processed a prompt?
These questions will become more urgent as Copilot moves deeper into Office documents, email, Teams meetings, Power Platform workflows, endpoint management, and security products. AI governance is not only about whether a model says something dangerous. It is also about whether an organization knows what system touched its data in the first place.
The AI Market Is Moving From Maximalism to Triage
SiliconANGLE’s report places Microsoft in a wider pattern: companies including Amazon, Accenture, Meta, and Uber have reportedly been looking for ways to reduce AI bills. That tracks with the broader shift from AI maximalism to AI triage.The early phase of generative AI rewarded brute force. Use the strongest model. Add more context. Generate more drafts. Turn on the agent. Let the model think longer. If the output was impressive, the cost could be rationalized as innovation.
That phase is ending. The next phase is about cost-aware intelligence. Enterprises want smaller models for narrow tasks, cheaper models for high-volume tasks, and frontier models only where the incremental capability changes the outcome.
This is not a retreat from AI. It is the industrialization of AI. Every major computing platform eventually learns the same lesson: performance matters, but performance per dollar matters more.
Microsoft’s Model Independence Is Also a Political Signal
There is a corporate politics layer here that should not be ignored. Microsoft’s relationship with OpenAI has been enormously productive, but also unusually complicated. Microsoft is investor, infrastructure provider, product partner, reseller, and partial competitor.Building MAI gives Microsoft a way to say, internally and externally, that it is not merely the cloud host for someone else’s brain. That matters to developers, customers, regulators, and Wall Street. It also matters to Microsoft’s own product groups, which need confidence that the company can ship AI features on its own schedule.
The reported use of MAI in Excel and Outlook is symbolically important because those are not side projects. They are core Microsoft products. If Microsoft is comfortable using its own models there, even for a small share of traffic, it is making a statement about trust.
The statement is not “we no longer need OpenAI.” It is “we need options.” In platform strategy, options are power.
The Bill for AI Is Finally Reaching the Product Roadmap
The concrete lesson from Microsoft’s reported shift is that AI costs are no longer a back-office concern. They are shaping which models users get, which features ship, and how vendors design their platforms.- Microsoft is reportedly routing some Excel, Outlook, and other productivity prompts to its own MAI models to reduce reliance on OpenAI and Anthropic.
- The shift appears limited for now, with MAI handling only a small fraction of Microsoft’s overall Copilot traffic.
- MAI-Thinking 1 is being positioned as a cost-efficient reasoning model rather than a universal replacement for the largest frontier systems.
- OpenAI remains central to Microsoft’s AI strategy, but Microsoft is clearly reducing the risk of depending on any single outside model provider.
- Anthropic’s premium pricing has become a visible pressure point as enterprises look for cheaper ways to run high-volume AI workloads.
- For IT departments, the next governance challenge is knowing which model processed which data, under which controls, and at what cost.
References
- Primary source: SiliconANGLE
Published: 2026-07-07T23:54:08.160912
Loading…
siliconangle.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 - Related coverage: thetechportal.com
Microsoft begins replacing OpenAI, Anthropic models with in-house MAI AI across key products: Report - The Tech Portal
Microsoft has reportedly begun deploying its own AI models across some of its most widely used productivity applications. Instead of
thetechportal.com
- Related coverage: hokai.io
Loading…
hokai.io - Related coverage: thenextweb.com
Microsoft AI chief wants to eliminate Anthropic spending
Mustafa Suleyman called Anthropic "extremely expensive" and unveiled seven in-house Microsoft AI models at Build to compete on price.thenextweb.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
- Related coverage: thurrott.com
Microsoft is Reportedly Transitioning to In-House AI Models - Thurrott.com
Microsoft has reportedly started transitioning away from OpenAI and Anthropic AI models in an attempt to cut costs.www.thurrott.com - Related coverage: santageai.com
Microsoft Launches Its Own AI Models at Build 2026, Cutting the OpenAI Cord | Santage
Microsoft unveiled seven in-house MAI models at Build 2026, headlined by MAI-Thinking-1 and MAI-Code-1-Flash. Project Polaris will replace GPT-4 Turbo in GitHub Copilot starting August 2026.santageai.com - Related coverage: officechai.com
Loading…
officechai.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 - Official source: microsoft.ai
MAI-Thinking-1 | Microsoft AI
microsoft.ai
- Related coverage: awesomeagents.ai
Loading…
awesomeagents.ai - Related coverage: itpro.com
Anthropic reveals Claude Opus 4.6, an enterprise-focused model with 1 million token context window for extended code capabilities | IT Pro
The AI developer highlighted financial and legal tasks, as well as agent tool use, as particular strengths for the new modelwww.itpro.com - Related coverage: axios.com
Microsoft debuts Scout agent, homegrown reasoning model
Microsoft is seeking to show it is a serious player in AI.www.axios.com
- Related coverage: techxplore.com
Loading…
techxplore.com