Microsoft Unveils Its Own Coding Model MAI-Code-1 at Build
On June 2, 2026, Microsoft unveiled its in-house coding models "MAI-Code-1" and the smaller "MAI-Code-1-Flash" at its Build 2026 developer conference, optimizing them for GitHub Copilot and beginning rollout. Part of a family of homegrown models aimed at reducing reliance on OpenAI and cutting costs, the originating @MicrosoftAI announcement points to the Build session "Microsoft AI coding models are optimized for GitHub Copilot."
According to the announcement, MAI-Code-1 and MAI-Code-1-Flash are becoming available in VS Code's GitHub Copilot (across all Free/Pro/Pro+/Max plans) and can be selected via the model picker or Auto picker. The smaller MAI-Code-1-Flash in particular is integrated and trained directly with Copilot's production harness (its tool-integration backbone), and is rolling out with no extra setup required. Official information is published on the Microsoft blog, the GitHub Changelog, and the Microsoft AI introduction page.
The two models are part of a newly introduced family of seven MAI models (including MAI-Thinking-1), and represent for Microsoft its "first in-house small coding models optimized specifically for Copilot." Until now Copilot has primarily used OpenAI's GPT series and others, but as CNBC reported, this push into homegrown models simultaneously targets reduced dependence, cost optimization, and Copilot-specific tuning. The announcement had been previewed in advance by Reuters and others.
According to its model card, MAI-Code-1-Flash has about 5B active parameters (some sources cite 137B total) and a 256K-token context length. Its training used "zero distillation" on clean, commercially licensed data, trained March–May 2026 and released June 2. Its standout trait is being trained directly on GitHub Copilot's production harness, optimizing it for agentic coding involving tool use and real-environment integration.
In Microsoft's own evaluations, the model reportedly scored 51.2% on SWE-Bench Pro (+16 points over Claude Haiku 4.5's 35.2%), 71.6 on SWE-Bench Verified (ahead of 66.6), +28.9 points on the instruction-following IF Bench, and 85.8% adjusted accuracy on an adversarial coding benchmark. It is also said to cut tokens by up to 60% on complex tasks, directly lowering costs. MAI models are slated to become available on external platforms such as Fireworks AI in the future.
Reactions spread on X immediately after the announcement, with welcoming voices highlighting "Microsoft-made rather than borrowed from OpenAI," "rollout to all tiers," and "cheaper and faster with 60% token reduction." An industry media post also emphasized advances in agentic coding. Meanwhile, because the presented benchmarks are largely Microsoft's internal evaluations, some take a cautious stance awaiting independent verification, and there were comments tying expectations for cost savings to Copilot's token billing. It is drawing attention as one of the highlights of Build 2026.