MAI Code
by Microsoft · Model · Language & code models
unproven
Microsoft says it nearly matches top coding models cheaply; no independent tests yet
MAI Code 2 for multi-step coding available in GitHub Copilot model picker, Copilot Fusion and Foundry
AccessHosted app
Why we say so
All the benchmark and cost claims, including 98% of GPT-6.1 Sol's score and a 64% blind preference over GPT-6 Sol, come from Microsoft, and the claims themselves concede it falls short of both rivals on the averages. Hacker News commenters question the cost comparison and the weight of launch benchmarks, and say they cannot find a way to try it outside GitHub Copilot. One commenter finds it about on par with Opus or Sol in daily work, which is a single voice.
Evidence
- Nov 18MAI Code 2 announced for enterprise coding; available in GitHub Copilot, with OpenRouter, Vercel, OpenCode, Kilo and Hermes planned
- Nov 18Claimed 98% of GPT-6.1 Sol's average score on DeepSWE, Terminal Bench 2.1, SWE-Atlas and FrontierCode at 40% of the cost
- Nov 18Claimed 96% of Claude Opus 5.5's average coding benchmark score at 5% of the cost
- Nov 18Not open weights, and no pricing page, API keys or playground found; users can't tell where to try it