Microsoft AI introduces MAI-Code-1-Flash for GitHub Copilot

Microsoft AI has introduced MAI-Code-1-Flash, a coding model built for fast, efficient assistance in everyday developer workflows. It is rolling out to GitHub Copilot individual users in Visual Studio Code, where it appears in the model picker and under the default Auto picker, with no additional setup required.

Trained on the production Copilot harness

Microsoft says the model was trained from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models. Training used the GitHub Copilot harnesses that run in production, so the model learns how to interact with surrounding tools and systems in agentic coding tasks. Checkpoints were evaluated on core software engineering tasks, repository question answering, refactoring and telemetry-grounded tasks adapted from real Copilot usage.

Benchmark results against Claude Haiku 4.5

The company reports higher pass rates on all four evaluations it ran in the same production harness: SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual and Terminal Bench 2. The widest gap is on SWE-Bench Pro, where MAI-Code-1-Flash scored 51.2% against 35.2% — a 16-point lead. On SWE-Bench Verified, Microsoft says the model solved harder problems with up to 60% fewer tokens.

On instruction following, it led by 28.9 points on IF Bench precise instruction following and by 14.5 points on the rubric-based Advanced IF. Microsoft adds that the model also outperforms Claude Haiku 4.5 on core reasoning in math, science and visual generation coding.

Adaptive thinking and value per token

MAI-Code-1-Flash was trained with adaptive solution length control, which adjusts the depth of a response to the task: concise answers for simple requests, more reasoning budget for complex work. Microsoft frames the result as value per token — fewer tokens mean lower latency and cost, and smoother interactive workflows.

Adversarial testing and availability

Microsoft also built a 186-question, 34-category benchmark around adversarial traps such as inverted classics, impossible tasks and underdetermined scenarios. MAI-Code-1-Flash reached 85.8% adjusted accuracy, with strong results in reasoning, instruction following and recognising impossible problems, although core adversarial categories such as Einstellung traps stayed below 50% accuracy. The model is available now to Copilot individual users in VS Code.