Microsoft has unveiled seven specialized in-house models designed to reduce the cost of agents that run continuously inside enterprise products. The lineup spans coding, cybersecurity, workflow decisions, image editing, speech and document retrieval, with several models already appearing in GitHub Copilot, Foundry and Microsoft security services.
Microsoft says MAI Code 2 averaged 98% of GPT-6.1 Sol’s performance across four coding benchmarks at 40% of the cost, using a temporary 50% price discount. The company also reports that MAI Cyber 2 Flash solved 92.3% of a cleaned CyberGym-E2E test set, slightly above GPT-6 Astra while using a smaller reasoning budget. These are vendor comparisons, and several tests ran on Microsoft’s internal harness.
Decision 1.1, built on Google’s Gemma 4 26B A4B, routes requests among developer-approved actions and can abstain when uncertain. MAI Realtime 1 handles speech directly and uses an interface compatible with the OpenAI Realtime API. A separate licensed-weights program will let selected customers fine-tune models inside Azure, but pricing and a firm launch date are pending.
The release gives Microsoft more control over costs and product integration while its non-exclusive OpenAI partnership continues through 2032.