What happened
Microsoft has introduced two coding models, MAI Code 2 and a lighter version called MAI Code 2 Flash, aimed at the everyday programming work that enterprises handle, Microsoft AI reported. The larger model is pitched at debugging, edits spanning many files, reasoning across whole repositories, interface building and shipping code to production. Flash targets high-volume jobs such as quick edits, routine fixes and code explanations.
According to the company, MAI Code 2 reached 98% of GPT 6.1 Sol's performance at 40% of its cost; against Claude Opus 5.5, it scored 96% as well while costing only 5% as much. Those figures are averaged across four benchmarks, SWE-Atlas, FrontierCode v1.1, DeepSWE 1.1 and Terminal Bench 2.1, and the costs use a 50% promotional price. In blind frontend tests judged by 258 annotators, reviewers favored its output in 64% of comparisons against GPT-6 Sol.
Microsoft said Flash performs comparably to Claude Haiku 5.5 and GPT-6 Luna on three routine-coding benchmarks, at 20% and 65% of their respective costs. That comparison also uses promotional pricing. Both models are available in GitHub Copilot. Flash is in a limited-access preview in Microsoft Foundry, with releases planned on Vercel, Kilo, Open Router, Hermes and Open Code. The company said neither model is yet tuned for fundamental science research or very long-horizon agentic work.
Context
Microsoft is also using the line inside its own products. It derived a single model, MAI Work Flash, from Code 2 Flash to serve several products across M365 and Fabric IQ. Internal results put it on par with GPT 6 Luna for Excel and PowerPoint tasks. In Word, it performed close to GPT 6.1 Sol while using about 42% fewer input tokens per call.
The briefing
The week in tech, in one email.
The stories worth your time, every week. Unsubscribe anytime.
