Microsoft's MAI-Code-1-Flash Is Landing in GitHub Copilot: Should You Switch?

For most of the last few years, much of the headline AI in Microsoft's own products, Copilot above all, leaned on OpenAI models underneath, an arrangement the company has been fairly open about. That left Microsoft in an unusual spot: a company whose products put AI in front of millions of people, yet that had not, strictly speaking, built the frontier models those products leaned on. At Build 2026 on June 2, that shifted in a concrete way: Microsoft's AI group introduced a family of models it trained itself, and one of them is already rolling out into the GitHub Copilot model picker.

The Short Answer

One of these models is something you can start using inside Copilot as the rollout reaches your account; the other is a benchmark headline that, outside a private preview, you cannot actually touch yet. One is news you can act on now. The other is a claim to keep at arm's length until someone outside Microsoft checks it. Keep them apart and the launch reads clearly, before the marketing settles over everything.

What Microsoft Announced at Build

Microsoft AI's Superintelligence team unveiled seven home-grown models at once. Two of them carry the weight of the announcement.

The flagship is MAI-Thinking-1, the group's first reasoning model. Microsoft describes it as a sparse mixture-of-experts design with about 35 billion active parameters out of roughly a trillion total, and a context window of 256,000 tokens. It is available now in private preview on Microsoft Foundry, with a public preview on the MAI Playground promised "soon."

The workhorse is MAI-Code-1-Flash, a small, fast coding model that is rolling out to individual GitHub Copilot users inside Visual Studio Code, where it appears as a choice in the model picker.

The rest of the family fills in the edges: MAI-Image-2.5 and a Flash variant for image generation, a transcription model (MAI-Transcribe-1.5) covering 43 languages, and a voice model (MAI-Voice-2, plus a Flash variant) that coverage says adds more than fifteen new languages. Useful, but supporting cast. The reasoning and coding models are where the strategy lives.

One detail Microsoft keeps repeating is how these were trained. The company says its reasoning model was built on "clean and appropriately licensed data," with AI-generated content excluded from pre-training and, pointedly, "without distillation from third party models." Translation: it did not learn by training on the outputs of another company's model. Hold onto that phrase. It is doing more work than it looks.

Here is the same launch in one view, with the part the marketing skips, that all of these figures are still Microsoft's own, kept in the last column:

The Model Arriving in Copilot

Skip past the keynote spectacle and here is the practical change. Microsoft is rolling MAI-Code-1-Flash out to individual Copilot users in VS Code, where it shows up as another entry in the model picker, alongside whatever other models your Copilot plan already offers. Depending on how far the rollout has reached your account, it may be there already or arrive soon after. When it does, switching is simple: open the model picker in Copilot inside VS Code and choose MAI-Code-1-Flash from the list.

It is built for a specific job rather than for topping every chart. "Flash" is the tell. This is a small model tuned for fast, lightweight, high volume coding work, the kind you want answering the twentieth small refactor of the day, not the one you reach for to architect a system. Microsoft says it trained the model directly against the same Copilot harnesses it runs in production, which is a sensible way to make a model good at the narrow thing it will actually be asked to do.

Microsoft's own numbers put the Flash model ahead of Anthropic's Claude Haiku 4.5, the comparison being another model aimed at the same small, fast coding niche. On the SWE-Bench Pro coding benchmark the company reports 51.2 percent for its model against 35.2 percent for Haiku, plus a claim that on SWE-Bench Verified, a separate test, it reaches its result using up to 60 percent fewer tokens. Fewer tokens is the metric Microsoft is leaning on, and it matters at scale: a model that finishes the same task with less text generated is cheaper to run, which is the point of a "Flash" tier. Whether that ever shows up on your own bill is a separate question, because it depends on how your Copilot plan meters usage, and Microsoft has not published pricing for this model. If you have been thinking about where AI coding costs actually accumulate, this is Microsoft aiming at that anxiety, even if the savings are still a number you would have to verify for yourself.

None of this makes it the model that wins demos. It looks built to take routine Copilot work off the heavier, slower options, assuming the quality holds up outside the company's own test sheet.

About Those Benchmarks

Now the flagship. MAI-Thinking-1 is the one Microsoft wants you to be impressed by, and the numbers it published are genuinely strong. The company reports 97.0 percent on the AIME 2025 math competition benchmark and 94.5 percent on AIME 2026, says the model goes "toe-to-toe with Claude Opus 4.6" on coding tasks, and points to a blind evaluation run by Surge, across 1,276 tasks, in which people preferred its model to Claude Sonnet 4.6.

Read those claims twice, because two things about them deserve a skeptical eye.

The plainest problem first: every number here comes from Microsoft. No independent results have been reported yet, and a vendor's own benchmark sheet is a marketing document as much as a technical one. That does not make the figures wrong. It makes them unconfirmed, and the gap between "the vendor says" and "outside researchers reproduced it" is exactly where AI launches tend to deflate.

There is a subtler thing going on, too. Notice who the model is measured against. Every headline comparison is to a Claude model, Opus 4.6 and Sonnet 4.6, and not to a GPT model from OpenAI. For a launch whose subtext is Microsoft loosening its dependence on OpenAI, the missing GPT comparison is conspicuous. The published numbers let Microsoft look competitive at the frontier, yet they leave untouched the one head-to-head many developers would most want to see. Whether that gap is deliberate or simply convenient is not something the launch materials answer, and it is worth holding the benchmarks a little more loosely because of it. If you have been following how the newer Claude Opus releases keep moving the coding baseline, you will see why standing level with even the 4.6-generation Opus is a result worth publishing.

There is also a quieter constraint: almost nobody can verify any of this yet, because the reasoning model is locked behind a private Foundry preview. The benchmarks are public; the model is not.

The OpenAI Subtext

Underneath the model specs is the move that actually matters, and Microsoft is not hiding it. Building a reasoning model "without distillation from third party models" is the company making the case, on its own benchmarks, which no outsider has yet checked, that it can reach the frontier on its own data rather than renting that capability from OpenAI indefinitely. Coverage of the launch framed it plainly as a move to reduce reliance on OpenAI and to lower costs for the developers building on Microsoft's platforms.

That matters beyond corporate maneuvering. Coverage of the launch from outlets like Thurrott and Neowin also reported that the models would reach beyond Microsoft's own catalog, with planned distribution through outside hosts like Fireworks AI, Baseten, and OpenRouter. A company that only wanted these models to power its own products would not bother shipping them where rival tools can reach them. The intent is optionality: give developers a Microsoft alternative at several points on the price and performance curve, so that no single outside lab is the only road to good results.

For a developer, the spreadsheet view is simpler than the politics. More credible models competing for the same job tends to push prices down and quality up. Whether MAI earns a permanent slot in anyone's stack is unknowable today, but the arrival of another serious contender is good for the people paying the bills.

Should You Switch Your Copilot Model?

The Flash model is the only piece you can act on now, so treat the rest as something to watch.

Try it if you run a lot of small, repetitive coding tasks through Copilot and you care about speed and a light footprint more than maximum reasoning depth: boilerplate, simple refactors, test scaffolding, quick explanations. That is the lane it was built for, and a fast, lightweight model that is "good enough" for routine work can be the right default even when a heavier model would technically write nicer code.

Stay on your current model if your Copilot work skews toward the hard problems: gnarly debugging, reasoning across many files, design changes at the architecture level. A small model is the wrong tool there, and Microsoft is not claiming otherwise. The model picker exists precisely so you can keep a stronger model one click away. If you are still deciding which coding assistant to standardize on in the first place, our rundown of the major terminal coding agents is a better starting point than any single new model.

If you want to decide rather than guess, give it a fair trial instead of a vibe check: run a small refactor, a test scaffold, a short code explanation, and one real bug fix through it, then the same four through whatever you use now, and compare the results side by side. A day of that tells you more than any launch chart.

Wait on the reasoning model. There is nothing to decide until it leaves private preview and someone outside Microsoft can benchmark it on tasks the company did not pick.

The Open Questions That Decide This

A launch this large leaves a few honest open questions, and the answers will arrive after the keynote, not from it.

Independent benchmarks are the big one. The moment MAI-Thinking-1 reaches public preview, outside evaluators will test it on suites Microsoft did not curate, and those results, not the launch slides, are what should move your opinion. Watch, too, for a head-to-head against a GPT model, which the launch materials did not include; that gap is worth noting on its own.

Day-to-day Copilot feel is the other. Benchmark percentages and the experience of coding with a model are not the same thing. A model can post a strong score on a test and still annoy you in the editor, or post a modest one and feel great. Because the Flash model is arriving in the Copilot model picker anyway, the cheapest experiment is to switch to it for a day of ordinary work and judge it yourself rather than from a chart.

And keep the cost claim in view. "Up to 60 percent fewer tokens" is a vendor figure with an "up to" in front of it. If lower spend is the reason you are interested, measure your own usage before and after instead of trusting the headline.

What Microsoft has really done is make a serious case that it can build frontier-class models without OpenAI's help, though on its own benchmarks that case still needs outside confirmation, and it backed the claim with one small model you can start using as it reaches your editor and one large model you mostly cannot use yet. One is a low-stakes thing to try as soon as it reaches you; the other only becomes a real story when the preview opens and independent numbers arrive.

Frequently Asked Questions

Can I use MAI-Code-1-Flash right now? If you use GitHub Copilot in Visual Studio Code, soon if not already. As of early June 2026, Microsoft is rolling it out to individual Copilot users as a selectable option in the model picker, so it may already be in your account or arrive shortly as the rollout completes. When it shows up, you switch to it the same way you would pick any other Copilot model.

Is MAI-Thinking-1 available to the public? Not yet for general use. It launched in private preview on Microsoft Foundry, with a public preview on the MAI Playground described only as coming "soon." Most developers will not be able to test it until that opens.

Does this mean Microsoft is dropping OpenAI? No. The launch shows Microsoft can build competitive models on its own and wants more options, but it does not announce an end to the OpenAI partnership. Read it as Microsoft reducing its dependence and widening developer choice, not cutting ties.

Are the benchmark numbers independently verified? No. Every figure cited here, the AIME scores, the coding results, the claim about fewer tokens, comes from Microsoft's own announcements. They are plausible and specific, but no outside lab has reproduced them yet, so treat them as vendor claims until that happens.

Source Links

• Microsoft AI: Introducing MAI-Thinking-1
• Microsoft AI: Introducing MAI-Code-1-Flash
• CNBC: Microsoft unveils new AI models to lessen reliance on OpenAI
• Thurrott: Build 2026, Microsoft launches its first flagship reasoning AI model and more
• Neowin: Microsoft unveils its MAI reasoning and coding models