Microsoft cuts MAI-Code-1.1-Flash price fourfold
Microsoft AI's updated in-house coding model costs a quarter as much as June's 1.0 release, uses fewer tokens per task, and now ships with downloadable weights for local use.
Microsoft AI announced MAI-Code-1.1-Flash, a point update to the small coding model it launched at Microsoft Build in June. The company says the new version is already in production in GitHub Copilot.
Microsoft says it focused on CLI tasks and .NET after hearing from developers. It reports a 22% improvement on Terminal-Bench 2.1 in GitHub Copilot CLI and a 15% improvement on .NET tasks. Microsoft also points to production metrics: code survival rose 4% and return visits rose 9%. The company did not publish absolute benchmark scores.
In Copilot, tokens stream 25% faster and the model uses 25% fewer tokens to complete a task. Microsoft prices 1.1 at one quarter of 1.0. It credits better training and serving efficiency, plus optimization across "more than hundreds of thousands" of reinforcement-learning environments in Copilot.
The model can also be downloaded and run locally. Microsoft quantized it to 3 bits, which lowers memory needs while keeping the full 256K context window. It says this version performs comparably to the full-precision model on SWE-Bench Verified and Terminal-Bench 2.1. Even so, the company recommends devices with more than 120GB of RAM. Local calls carry no inference charges.
The on-device model gives Copilot's router a local option: eligible coding work runs on the user's hardware, and harder tasks go to cloud models. Experimental access is due by the end of the month in the GitHub Copilot app, Copilot CLI, and Visual Studio Code. Microsoft did not disclose parameter counts or license terms for the weights.