Microsoft's new code model MAI-Code-1.1-Flash is cheaper than its predecessor, but weaker and more expensive than DeepSeek

Microsoft published MAI-Code-1.1-Flash, a code model for GitHub Copilot. According to Microsoft, it delivers better code with 25% higher token efficiency, costs a quarter of its June predecessor, and developers kept 4% more of its code. It was trained with hundreds of thousands of reinforcement learning environments in GitHub Copilot. In benchmarks, the model is slightly ahead of its predecessor and mini models from Anthropic and OpenAI, but significantly behind DeepSeek-V4-Flash-0731. For example, on Terminal-Bench 2.1, MAI-Code-1.1-Flash scored 62.9% compared to DeepSeek's 82.7%. Pricing is also higher: input at $0.20 vs $0.14, output at $1.20 vs $0.28 per token. Microsoft hides benchmark results in the model card and uses vague metrics in its announcement. This conflicts with Microsoft's narrative as an advocate for open AI models, as it invests in a proprietary, weaker, and more expensive model instead of using available open-weight alternatives like DeepSeek. The likely reason is optimization of its own margins, aiming to establish its models as the standard within its ecosystem.