Microsoft launched the MAI Code 1.1 Flash code model for GitHub Copilot, claiming it writes better code, is 25% more token-efficient, and costs a quarter of the previous version's price. The Decoder pointed out, however, that the model is more expensive and underperforms the DeepSeek-V4-Flash-0731, both in price and in benchmarks.
According to Microsoft, the new model had code acceptance increase by 4%, and training involved 'hundreds of thousands of reinforcement learning environments in GitHub Copilot.' The company did not provide direct comparisons with competitors in the announcement, limiting itself to metrics such as 'code survival rose 4%' and 'return visits increased 9%'.
Performance
In benchmarks, the MAI Code 1.1 Flash scored 72.6% on SWE-bench Verified, surpassing its predecessor (71.6%), Claude Haiku 4.5 (69.8%), and the GPT-5.4 mini (69.2%). DeepSeek-V4-Flash-0731 did not disclose a result on that test. On Terminal Bench 2.1, however, Microsoft's model scored 62.9%, below DeepSeek's 82.7%, but above the other Anthropic and OpenAI models.
The price comparison per million tokens shows DeepSeek-V4-Flash with input at US$ 0.14 and output at US$ 0.28. MAI Code 1.1 Flash costs US$ 0.20 and US$ 1.20, respectively. Claude Haiku 4.5 has higher prices: US$ 1.00 and US$ 5.00.
Strategy
The analysis points out that the launch contrasts with Microsoft's rhetoric in favor of open AI, since the company is investing in a proprietary model that is more expensive and has lower performance than available alternatives. The company would have replaced OpenAI and Anthropic models in Copilot with its own MAI models to reduce costs, and the trend is for its models to become the standard in the ecosystem, limiting users' options.
