Overview
- MAI Code 2 and MAI Code 2 Flash, which Microsoft released Tuesday, are available in GitHub Copilot and the Flash variant is in a limited Microsoft Foundry preview.
- Microsoft says MAI Code 2 matches about 98% of GPT‑6.1 Sol’s coding performance at roughly 40% of the cost and about 96% of Claude Opus 5.5’s at about 5% of the cost, and that MAI Cyber 2 Flash scored 92.3% on CyberGym‑E2E versus GPT‑6 Astra’s 91.5% in its tests.
- Those performance and price claims come from Microsoft and partner evaluations that use promotional pricing, selective test sets and task exclusions, so independent benchmarks and real‑world trials are needed before generalizing the results.
- Microsoft will embed MAI models across its stack — GitHub Copilot, Foundry and Azure — and offers licensed model weights, on‑Azure hosting and its Maia 200 accelerator to control costs and let customers customize deployments.
- For developers and security teams the practical tradeoff is routing repeatable, high‑volume tasks to cheaper specialist models while keeping larger generalist models for hard problems and retaining human review; Microsoft plans to start rolling MAI Cyber 2 Flash into its MDASH security products in mid‑December.