MAI-Code-1-Flash
Microsoft releases MAI-Code-1-Flash, a 5B-parameter coding model that outperforms Claude Haiku 4.5 on benchmarks while using fewer tokens.
MAI-Code-1-Flash is a new coding model built by Microsoft and integrated into GitHub Copilot for VS Code. The model uses adaptive solution length control to adjust response depth based on task complexity, solving harder problems with up to 60% fewer tokens than competitors.
On SWE-Bench Pro, MAI-Code-1-Flash achieved 51.2% pass rate versus Claude Haiku 4.5's 35.2%, a 16-point lead. It also outperformed on SWE-Bench Verified, SWE-Bench Multilingual, and Terminal Bench 2, as well as instruction-following and reasoning benchmarks. The model was trained directly with GitHub Copilot harnesses to optimize for real production workflows rather than benchmarks alone.
What HN community is saying
Commenters raised concerns about benchmark methodology, noting Microsoft's framing as "hill climbing" raises questions about potential train-test contamination, though Microsoft provided decontamination details in supplementary materials. Several threads discuss the gap between benchmark scores and practical usability: a 51% pass rate on SWE-Bench Pro is on par with Claude Opus 4.6 but insufficient for fully autonomous code generation without human verification.
Other discussion focused on the model's non-open-weight status, contrasting with Microsoft's historical Phi releases. Commenters noted the broader trend of small specialized models for local inference and pricing: MAI-Code-1-Flash costs $0.75 per 1K input tokens and $4.50 per 1K output tokens via GitHub Copilot.