MAI-Code-1.1-Flash

MAI-Code-1.1-Flash is Microsoft's second in-house coding model, released August 11, 2026, ten weeks after MAI-Code-1-Flash debuted at Build, and the headline is price. GitHub lists it at $0.20 per million input tokens, $0.02 cached, and $1.20 output, exactly a quarter of the $0.75, $0.075, and $4.50 that 1.0 charged. The model card describes a sparse mixture-of-experts with 138 billion total parameters and 5 billion active, a 256K token context window, and text plus image input, which 1.0 did not take. It started from the compressed 5B-active mid-training checkpoint of MAI-Thinking-1 rather than a fresh pretrain. Microsoft reports a 22 percent improvement on Terminal-Bench 2.1 inside GitHub Copilot CLI, a 15 percent gain on .NET tasks, 25 percent fewer tokens per completed task, 25 percent faster streaming, and code survival up 4 percent in production, all measured on its own harness and none independently reproduced. Weights are closed, and the license is whatever product terms apply where it is deployed, which in practice means GitHub Copilot first. GitHub deprecated MAI-Code-1-Flash on September 10, 2026, so this is now the Microsoft coding model to budget against Claude Haiku 4.5 and the Gemini Flash tier.