Microsoft Builds Its Own AI Models — and Wants Off OpenAI

Redaktion · · 5 Min. Lesezeit

At its Build developer conference (June 2–3, 2026), Microsoft unveiled its own family of AI models: the MAI line, seven models including the reasoning model MAI-Thinking-1 and the coding model MAI-Code-1-Flash. The stated goal is to reduce reliance on external providers like OpenAI and Anthropic — while lowering costs for developers using GitHub Copilot. MAI-Code-1-Flash has been available in the Copilot model picker since June 2, 2026 and is rolling out gradually across all Copilot plans.

What used to be the case

GitHub Copilot ran for years on OpenAI models, later joined by Anthropic. That made Microsoft a reseller of third-party model performance: every Copilot request created costs at an external provider, which Microsoft absorbed. Strategically uncomfortable — especially after Microsoft renegotiated its OpenAI partnership in late 2025.

At the same time, Microsoft switched Copilot’s billing to a token-based model on June 1, 2026: instead of a flat monthly fee, users now pay by tokens consumed, billed via “GitHub AI Credits”. That drew noticeable pushback — in cases cited by TechCrunch, individual developers reported bills jumping from around $29 to several hundred dollars a month. The complaint: Microsoft had encouraged heavy usage and was now sending the bill retroactively.

What’s true now

1. Microsoft has its own frontier models. With MAI-Thinking-1, Microsoft owns its first reasoning model; with MAI-Code-1 and MAI-Code-1-Flash, its own coding models. In blind tests cited by Microsoft, MAI-Thinking-1 was reportedly favored over Anthropic’s Claude Sonnet 4.6 and matched Claude Opus 4.6 on the SWE Bench Pro coding benchmark. These figures come from Microsoft itself and are not yet independently verified.

2. MAI-Code-1-Flash runs right inside Copilot. The model wasn’t just tested against the Copilot environment — it was trained inside its production harness. Since June 2, 2026 it’s in the VS Code model picker. Per the GitHub changelog it delivers “best-in-class quality for its size” and suits lightweight coding workflows — the changelog gives no precise benchmark figures.

3. The cost advantage is the real point. Several reports cite a token price for MAI-Code-1-Flash of roughly $0.75 input / $4.50 output per million tokens — versus $5 / $30 for GPT-5.5 under Copilot. They also say the model uses about 60 percent fewer tokens on hard tasks. Both numbers come from secondary sources, not the official changelog — but as an order of magnitude they’re consistent with Microsoft’s cost argument.

Reading it right

The headline “Microsoft builds its own models” is less a model event than an economic one. Microsoft closes a loop it has signaled since the OpenAI renegotiation in late 2025: its own reasoning model, its own production-grade coding model — and thus a credible path off OpenAI’s bill. Whoever owns the infrastructure (Azure) and the model controls the margin.

The timing is no accident. The token-based Copilot billing from June 1 had just made the cost question painfully visible. A homegrown, markedly cheaper model in the same picker is the obvious answer: Microsoft can offer developers a cheap default option without paying OpenAI or Anthropic per token. For users that can mean real savings — provided MAI-Code-1-Flash’s quality carries the routine work it’s meant for.

The necessary caution: all performance and efficiency numbers so far come from Microsoft or secondary reports. “Favored in blind tests” and “60 percent fewer tokens” are vendor or reporting claims, not independently checked facts. And a cheap model helps little if it can’t keep up on the genuinely expensive, complex tasks — exactly where the token consumption comes from.

What you can do now

If you use GitHub Copilot: Try MAI-Code-1-Flash for lightweight tasks (boilerplate, small refactors, tests) and compare credit consumption against GPT-5.5 or Claude on your real tasks. With the new token-based billing, model choice now hits the bill directly.

If the Copilot change hit you: Measure your actual token consumption per typical task before judging the plan. The reported extreme bills mostly came from heavy, iterative workflows — a cheaper default model can meaningfully cut consumption.

If you set provider strategy for clients: Treat MAI as a signal, not a finished recommendation. The coding-model market keeps fragmenting — alongside OpenAI and Anthropic, Microsoft is now in the race with its own model. Lock-in to a single provider gets more expensive to reverse the more options emerge.

Entdecke mehr