November 18, 2026CodingAgentsInfrastructure

MAI Code 2: Microsoft Builds Its Own Cheap Workhorse for Copilot

Microsoft used Ignite day one to put its own models under GitHub Copilot, on the same afternoon OpenAI shipped GPT-6.2. MAI Code 2 and MAI Code 2 Flash are coding models built for what Microsoft calls everyday enterprise coding, and the company says plainly they are not yet tuned for fundamental science or very long-horizon agentic work. Announcement: https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/

The claims, all Microsoft's: averaged across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, MAI Code 2 reaches 98% of GPT-6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5 at 5% of the cost. In blind frontend evals with 258 annotators it was preferred over GPT-6 Sol 64% of the time. Code 2 Flash is pitched as roughly GPT-6 Luna or Claude Haiku 5.5 at 65% and 20% of their cost, though the Flash cost in that chart is a 50% promotional price. The evals ran on Microsoft's MSBench through GitHub Copilot CLI with no subagents. Both are in the Copilot model picker now, Flash is in limited preview on Foundry, and OpenRouter, Vercel, OpenCode and Kilo are listed as coming soon.

The more interesting launch is MAI Licensed Weights: https://microsoft.ai/news/introducing-mai-licensed-weights/ . Trusted customers get MAI weights, tune them on their own data and run them inside their own Azure tenant. Microsoft's case studies: AT&T at 1/50th the cost of GPT-6.1 Sol on standard tasks, Woven by Toyota getting 44% ready-to-merge fixes on MISRA rules where frontier models needed a human, Alaska Airlines going from 81.2% to 91.1%. It is coming soon, through what Microsoft calls Frontier Company engagements, so for now it is a sales motion, not a download.

The rest of the keynote filled in around it: MAI Cyber 2 Flash for cyber defense, Copilot Workspaces where people and their agents share context, and Agent 365 Business Analytics, which Microsoft describes as an ERP for AI usage, tracking spend against outcomes across what it says are more than 50 million agents. Keynote post: https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/

Read the day as Microsoft hedging its biggest partner in public. The model is good enough at 40% of the price, the weights live in your tenant, and the bill gets its own dashboard. That is the cheap-workhorse half of the manager-and-worker setup, sold by the company that owns the editor.
← Previous
GPT-6.2: Cheaper, Smarter, and Quieter About What Broke
Next β†’
Gemini Tasks Goes Free, and Your Temporary Chats Get Longer Memory
← Back to all articles

Comments

Loading...
>_