Microsoft Copilot App Rebuild: Autopilot Agent Preview, Code Credits and Admin Costs — Megathread

The three stories point the same way. AI vendors are trying to make models that do work, not just chat. That raises practical questions about cost, permissions, reliability and who controls the data.

Microsoft's new Copilot: Home, Code and Autopilot​

This item has the strongest confirmation. In an Official Microsoft Blog post on September 25, Jared Spataro, Microsoft's chief marketing officer for AI at Work, described three new parts of the Copilot app:

• Home: the new starting point, where Chat and Cowork come together, with Word, Excel and PowerPoint built into the experience through Office in Copilot
• Code: lets everyone build their own solutions with the tools to run them safely, and is powered by the same underlying technology as GitHub Copilot
• Autopilot: a persistent, proactive and personal agent that keeps working even when you're not

DigitalToday said Scout, first shown at Build this year, has been renamed Autopilot. Microsoft confirms it: "Autopilot, previously called Scout, is your digital teammate." According to Microsoft, you give it a name, a role and a goal, and it goes to work — watching channels, following up on threads, running recurring work and picking a project back up days later, without waiting for a prompt.

Availability​

This is not a feature you can switch on everywhere today. Microsoft says Home and Code will start rolling out in its Frontier program in the coming weeks and Autopilot is expanding to private preview at the end of the month. Super Simple 365, a Microsoft 365 explainer site, notes that a private preview is a limited trial for organisations Microsoft selects, so most people won't see it working yet. It adds that Microsoft says it will share rollout timing and admin controls later.

Code needs particular attention from admins. Super Simple 365 says Code uses Copilot Credits, you need a Microsoft 365 Copilot licence, and your organisation needs a spending policy in place. Microsoft's blog adds that Code will reach Microsoft 365 Premium and Pro subscribers in preview later this year.

Admin impact: billing and runtime​

The biggest change for IT may be billing, not the interface. Microsoft now splits Copilot spending in two:

Microsoft is also launching Copilot Managed Runtime, hosting infrastructure that lets code run safely right inside your company's Microsoft 365 environment. Microsoft says it is governed by IT but easy for everyone else.

If you run Power Platform, you will probably ask whether Code replaces it. Empowering.cloud quotes Microsoft's answer: "Copilot Code does not replace Power Apps or Power Automate... they begin from different development models and support different workflows." The same analysis treats Microsoft's promise of automatic routing between Chat, Cowork and Code as a promise, not a feature until it ships. That is a fair stance.

Section summary: The Copilot overhaul is confirmed, but most of it is in Frontier or private preview. Agent features are billed by usage, so set spending policies before you enable them.

TypeSafe AI's Jev: a model that decides instead of writing​

DigitalToday's main story is Jev, from TypeSafe AI. DigitalToday describes the company as founded by a former OpenAI employee. The article says experts see Jev as structurally different from large language models: it doesn't generate text, and it is built to make decisions inside software. The article also cites a report that TypeSafe wants to raise $1 billion, and a linked headline raises a possible $10 billion valuation. No primary source confirms either figure, so treat them as reports.

TypeSafe announced Jev on September 15 in a post by founder Diogo Almeida. He writes that at OpenAI he helped develop the instruction-following methods that became the research behind ChatGPT. That is the company's own account.

TypeSafe calls Jev its first "System One Model," named after Daniel Kahneman's fast, intuitive System 1 thinking. The company's short description: unstructured input goes in, typed decisions with probabilities come out. A developer defines the possible answers in advance, such as a category, a score or a true/false flag. Jev returns one of them with a confidence estimate. Software can use that result directly, without parsing free-form text.

TypeSafe's published claims include:

• Pricing: input tokens at $0.042 per million, with output free
• Latency: 70 to 500 milliseconds end to end
• Training: a new method it calls Reinforcement Learning for Calibrated Decisions (RLCD)
• Limits: choices of up to 255 options; for larger sets, the company uses a two-stage process

Why developers care​

Vercel, which offers Jev through its AI Gateway, lists practical uses. They include choosing an agent's next tool, deciding whether a workflow should continue, retry, ask the user or stop, scoring risk before an action, and sending uncertain cases to a person for review.

Take a help desk or document pipeline that currently asks a general-purpose model to "classify this ticket" and then cleans up the text it returns. A model that only returns a valid label and a confidence score could make that step faster and more predictable.

Vercel also reports strong early uptake. It says Jev reached more than twice as many paid teams in its first 24 hours as any earlier model launch on the gateway. Nearly 13% of paid teams were using it by hour 24. Those figures cover only Vercel's gateway, and Vercel itself says the next test is whether the early adoption lasts.

Read the claims carefully​

TypeSafe says Jev "can't hallucinate." That claim is narrower than it sounds. It means the model can't return output outside the schema you define. It does not mean the chosen answer is correct: a well-formatted wrong answer is still wrong.

The company is fairly open about the limits of its own benchmarks. It says its workflow tests were built by its own team, so "some bias could exist." It says its speed figures were measured from the US West Coast, where the service is based. It also concedes it cannot prove its pricing isn't subsidised. The headline claims of 193.6 times faster and 444.6 times cheaper apply to what TypeSafe calls "System One" tasks, and the company expects those numbers to be at the high end of real-world gains.

Section summary: Jev is a narrow tool for routing and scoring decisions inside software, not a ChatGPT replacement. The performance numbers are the vendor's own, and the adoption data comes from one gateway on day one.

Meta's Muse and the personal-agent race​

Meta announced Muse on September 8, calling it a personal AI agent built for everyone. DigitalToday says Muse has made an impressive start. One headline it links claims Muse topped the App Store ahead of ChatGPT, but the roundup gives no date, country or source for that ranking. "Topped the chart for a while in one store" is a long way from "beat ChatGPT."

DigitalToday also reports that OpenAI plans to enter the personal-agent race. It says OpenAI hired developers of the open-source agent tool OpenClaw earlier this year to build a next-generation personal agent. One linked headline sums up the stakes: competition among personal agents is a battle for personal data.

That is also the question behind Microsoft's Autopilot. A consumer agent that books travel and an enterprise agent that follows up in Teams threads both need broad access to act on your behalf. Microsoft says Autopilot runs inside the customer's tenant with its own identity, audit and governance. Those promises need to hold up once admins test the controls.

The rest of the roundup​

DigitalToday also lists these developments. They are reported by the outlet and were not checked individually:

• Cheaper models: OpenAI and Anthropic released lower-cost models, reportedly to meet demand for lower AI spending and to compete with cheaper open-weight models.
• ChatGPT voice agents: the mobile app can now draft documents and emails and summarise Slack messages by voice.
• AWS Strands Harness: an open-source tool for building and deploying agents across environments.
• Google: Gemini 4 is reportedly coming earlier than its end-of-year target, and YouTube has a tool for editing videos with natural-language instructions.
• Akamai and Anthropic: a reported $11.6 billion, seven-year computing-capacity deal.
• Databricks: acquired spreadsheet startup Row Zero, with plans to use spreadsheets as an interface linking AI agents and business intelligence.
• DeepSeek: CEO Liang Wenfeng reportedly told investors Huawei could start supplying training chips as early as the fourth quarter.
• South Korea:
• Douzone Bizon set up an AI Governance Office based on ISO/IEC 42001.
• Wrtn Technologies picked underwriters for a planned IPO.
• Xenon won preliminary approval for a KOSDAQ listing.
• RealWorld signed an agreement with CJ Logistics to develop a logistics-focused robotics model.
• Consumer spending on generative AI is projected to pass 1 trillion won this year.
• Software industry: established vendors are offering discounts and free AI features to counter OpenAI and Anthropic, and more startups are building their own models.

What Windows and Microsoft 365 teams should do​

This is my recommendation, based on the announcements above and general industry experience:

• Check your Frontier enrolment if you want early access to Home and Code, and decide who in your organisation gets to try them first.
• Set spending policies before you enable Code. Agent work is billed by usage, so a busy agent can run up costs quickly.
• Wait for Microsoft's admin controls for Autopilot, including permissions, audit and memory, before you allow an always-on agent near sensitive channels.
• Test decision models like Jev on small, reversible steps first, such as tagging or routing, and send low-confidence results to a person.

The big question for the rest of 2026 is less about which model is smartest and more about which agent you trust to keep working unsupervised. The early announcements give only part of the answer.

Update: Additional details (September 28, 2026)​

UC Today reports that each Copilot Autopilot instance will have its own cloud computer, workspace, storage and identity. The outlet, citing The Decoder, says users can start an Autopilot job by @mentioning it in Teams, Outlook or a document, after which it continues running in the cloud while they are offline.

Microsoft has not yet published Autopilot-specific admin guidance showing whether its private-preview agent will use the same tenant approval, audience-scoping and permission-review controls available for custom Microsoft 365 agents.

Update: Additional details (September 30, 2026)​

Microsoft says a new plugin registry is rolling out now, with broader availability across Copilot surfaces planned in the coming weeks. The registry is intended to give IT one catalogue for Microsoft, partner and custom plugins, with central approval and management. Microsoft also says Fabric IQ is generally available in Copilot Chat and Cowork, while Code integration is coming through Frontier; Dynamics 365 and Power Platform grounding is due in public preview over the next month.

The announcement adds further cost-control specifics. Admins can set budgets, spending limits, policies and permitted model families for user groups, while users can view their remaining credits in Copilot. Microsoft says Agent 365 cost management is expanding to Code and Copilot Managed Runtime, with Copilot Studio agent coverage planned for October.

Update: Report names OpenAI’s Dots rival and quantifies Muse’s early uptake (October 4, 2026)​

OpenAI has reportedly moved from planning a personal agent to launching one called Dots, according to Foreign Policy Journal’s October 4 report on a Truist Securities analysis. The outlet says Truist compared Dots directly with Meta’s Muse and cited 2.8 million early Muse downloads. That adds a concrete adoption figure, but does not establish active usage, retention or the previously reported App Store ranking.

According to Foreign Policy Journal, Truist sees Muse’s advantage primarily in Meta’s distribution rather than superior AI capability. The report also describes Muse for Small Business as able to review sales, marketing and expenses while requiring owner approval before publishing, sending messages or spending money. That reported approval boundary is relevant for businesses considering delegated work.

For Microsoft 365 teams, Foreign Policy Journal reports that Dots has its own cloud computer and browser, alongside Slack, Microsoft Teams access and Microsoft security integration. The report does not specify supported Microsoft security controls, licensing or tenant-administration requirements, so those integration claims should not yet be treated as evidence of deployment readiness.