Microsoft launches MAI-Code-1-Flash coding model to challenge Anthropic’s Claude Haiku

Microsoft has introduced MAI-Code-1-Flash, a new artificial intelligence coding model designed to streamline developer workflows through improved speed and efficiency. The 5-billion-parameter model is being rolled out to individual GitHub Copilot users within Visual Studio Code, accessible via the model picker and the default auto picker. Microsoft states the model was built end-to-end using clean, appropriately licensed data and trained directly with production harnesses from GitHub Copilot to optimise performance in real-world agentic coding tasks.

The company claims MAI-Code-1-Flash outperforms Anthropic’s Claude Haiku 4.5 across several coding benchmarks, including SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual, and Terminal Bench 2. On SWE-Bench Pro, a metric for evaluating real-world software engineering tasks, the new model achieved a 51.2 per cent pass rate compared to Claude Haiku 4.5’s 35.2 per cent. Microsoft also reported that the model uses up to 60 per cent fewer tokens to solve complex problems on SWE-Bench Verified, suggesting a reduction in latency and cost for interactive workflows.

MAI-Code-1-Flash features adaptive solution length control, allowing it to adjust the depth of its response based on task complexity. This mechanism enables the model to remain concise for simpler requests while allocating more reasoning budget to problems requiring deeper analysis. Microsoft notes that this alignment between training, evaluation, and production environments is intended to ensure that offline improvements translate into tangible quality gains for developers using the tool daily.

In addition to coding benchmarks, Microsoft evaluated the model’s instruction-following and reasoning capabilities. MAI-Code-1-Flash demonstrated a 28.9 point lead on IF Bench for precise instruction following and a 14.5 point lead on rubric-based Advanced IF compared to its competitor. In a custom 186-question adversarial benchmark designed to test reasoning over pattern-matching, the model reached 85.8 per cent adjusted accuracy, surpassing Claude Haiku 4.5.

The release comes as Microsoft’s Microsoft AI (MAI) lab continues to expand its infrastructure and talent. The lab operates an operational GB200 cluster and is currently hiring for next-generation model development. Microsoft emphasised that the model supports agentic execution for multi-step workflows and is optimised for native integration into developer environments, aiming to reduce debugging time and improve overall coding efficiency.