Microsoft is ending its Claude Code licenses and moving developers onto a homegrown coding model built for GitHub Copilot. By June 30 the company’s own engineers have to drop Anthropic’s tool, and at the Build conference opening in San Francisco this week Microsoft is unveiling Project Polaris, the in-house model meant to anchor Copilot. The wager underneath it: distribution beats the best model.
Microsoft lost on model quality, and it is not pretending otherwise. Rather than put the better-rated tool inside its own product, the company is shipping a model it controls and leaning on the rails it already owns, GitHub, Azure, Windows and Office, to push a good-enough assistant into enterprises that would sooner flip a switch than run a procurement fight.
The License Microsoft Pulled From Its Own Engineers
In December, Microsoft handed thousands of its engineers Claude Code on the company dime. Inside its Experiences and Devices group, the division behind Windows, Office, Outlook and Teams, it quickly became the preferred coding tool. Then the invoice landed.
Claude Code is priced by the token. Engineers used it hard, and the bill blew past the division’s annual AI budget months ahead of schedule. The tool developers reached for first turned out to be the one whose cost Microsoft could not predict. By June 30, those teams have to stop using it and move to the GitHub Copilot CLI (command-line interface), the terminal version of Microsoft’s own assistant.
Anthropic, the AI lab founded by former OpenAI staff, released Claude Code in May 2025 and rode it to Claude Code’s billion-dollar revenue milestone inside a year, much of it from exactly this kind of heavy enterprise use. The same lab sells Anthropic’s paid Claude tier to individual developers, the cohort that took to the coding agent fastest. So the optics here are hard to miss: Microsoft owns GitHub, gave its engineers a rival’s agent, watched them prefer it, and is now taking it back.
Cost pulled the trigger, but cost is not the whole story. Renting intelligence by the token is a price Microsoft does not set, and depending on a competitor’s tool is an exposure it would rather not carry. Building its own model answers both at once, which is the same move the company is making in public at Build.
Project Polaris and the Drive to Drop OpenAI
The coding model is one piece of a wider campaign. Under Mustafa Suleyman, the DeepMind co-founder who runs Microsoft AI, the division has spent the past year chasing what it calls self-sufficiency, building a family of in-house MAI models so the company depends less on OpenAI for the intelligence inside its products.
The Coding Model at Copilot’s Core
Project Polaris, the code-named coding system, is slated to replace the OpenAI model at the heart of GitHub Copilot by August. For two years Copilot was the default AI coding assistant, bundled into the platform most developers already use, until Claude Code overtook it on quality. Microsoft’s answer is not a better model bought from elsewhere. It is one tuned on the company’s own repositories and run on its own cloud, part of a broader push toward building each model end to end.
A Full MAI Lineup
Polaris arrives alongside new MAI models for voice, image and transcription, first shown in April on the MAI models inside Microsoft Foundry. The MAI Superintelligence team, formed in November 2025, says its transcription model already beats OpenAI’s Whisper on accuracy while running on far less hardware.
- By August, Microsoft’s coding model is set to replace the OpenAI model inside GitHub Copilot.
- 3.8% word error rate for MAI-Transcribe-1 on the FLEURS multilingual speech benchmark, ahead of OpenAI’s Whisper.
- Under one second for MAI-Voice-1 to generate 60 seconds of speech on a single GPU (graphics processing unit).
We’re able to deliver the model with half the GPUs of the state-of-the-art competition.
That is Mustafa Suleyman, chief executive of Microsoft AI, describing the transcription model to VentureBeat. The economics are the point. A model Microsoft owns and runs cheaply is a cost it controls, unlike a per-token bill set by a rival.
The Survey That Explains the Bet
The numbers behind the move come from The Pragmatic Engineer survey of roughly 15,000 developers in February 2026. 46% named Claude Code their most-loved tool, more than double Cursor’s share and over five times GitHub Copilot’s. On its own, that reads like a rout.
Look closer and it inverts. Preference flips with company size: at organizations with more than 10,000 employees, GitHub Copilot was the most commonly used tool, while at smaller firms Claude Code ran away with it. The best-loved product wins where individuals pick their own tools, and the incumbent holds the accounts where committees do the buying.
| AI coding tool | Most-loved share | Where it wins |
|---|---|---|
| Claude Code (Anthropic) | 46% | Small companies, complex agentic tasks |
| Cursor | 19% | Startups, in-editor editing |
| GitHub Copilot (Microsoft) | 9% | Enterprises with 10,000+ staff, autocomplete |
Why Distribution Compounds and a Model Lead Fades
The split is structural, not a quirk of branding. It tracks who already owns the place where the work happens.
Owning the Rails
Microsoft owns GitHub, the development environments, the enterprise agreements and the cloud the code already runs on. When an AI feature lives inside tools a company has standardized on, switching it on across thousands of engineers is a configuration change, not a purchase. A better standalone product has a harder road, and the friction stacks up fast. Building each model itself, the core of Microsoft’s in-house AI model strategy, lets the company push that feature through channels it already controls.
- A rival tool has to be purchased on a new contract, not added to an existing one.
- It has to pass security and compliance review as a fresh vendor.
- It has to be wired into pipelines the incumbent already touches.
- Its per-token cost has to be justified line by line, the same math that sank Claude Code inside Microsoft.
This is the playbook that let Microsoft bundle Teams into Office and walk past Slack and Zoom, a comparison the survey’s own authors reach for.
Why the Best Model Keeps Getting Matched
The frontier models are excellent and close together. Claude, GPT and Gemini trade the lead every few weeks; a model that is best in May is matched in June. When the top tools are all good and only weeks apart, the edge shifts from owning the best model to owning the path customers reach it through. A model lead depreciates by the month, while distribution keeps compounding.
Who Else Owns the Rails
The same logic favors anyone who already controls the distribution. Microsoft is the cleanest case, with GitHub, Windows, Office and Azure feeding into one another.
Google can route AI straight into surfaces it owns outright, from Chrome to Android to Workspace, the same move on display as Google folds AI into the search surfaces it controls. The big enterprise-software vendors do a version of the same thing, carrying AI into accounts through systems their customers already run rather than asking them to adopt something new.
In each case the durable asset is identical, and it is not the model. It is the installed base, the layer nobody applauds, that turns a good-enough tool into a default, the kind of advantage visible in GitHub Copilot’s steady push into Visual Studio.
What the Bet Hasn’t Settled
There is a real caution buried in the move. The market keeps paying a premium for model quality as if the best model were the prize, and the coding-tool fight is the live counterexample.
The test is narrow and specific. Microsoft’s good-enough coding model has to hold enterprise seats even while developers keep naming Claude Code their favorite. The tool they rate highest is the one losing the enterprise, and it is losing to a company that concedes the model and owns the channel.
The switch inside Microsoft’s own walls is the first data point. The company is forcing its engineers off the tool they liked and onto the one it owns, and watching whether the work still gets done. Whether outside customers stay put through a full renewal cycle is unproven.
If Polaris holds those seats after it replaces the OpenAI model in Copilot this August, Microsoft will have shown that distribution beat quality out in the open, and that becomes the template for every layer of the stack above the chips. If enterprises drift back toward the tool their developers actually want, the premium on model quality was real all along.
