Microsoft announced its MAI family of fully in-house artificial intelligence models at the Build 2026 developer conference held in San Francisco on June 2. At the heart of the lineup—the company's first major foundation models built without OpenAI technology—are the coding-focused MAI-Code-1-Flash and the reasoning model MAI-Thinking-1. Microsoft said the move is intended to lower developer costs and reduce its dependence on outside providers.
MAI-Code-1-Flash is a compact, fast 5-billion-parameter coding model that became available across all GitHub Copilot subscription tiers starting June 2. According to figures shared by Microsoft, the model holds a 16-point lead over Anthropic's Claude Haiku 4.5 on the SWE-Bench Pro benchmark (51.2% versus 35.2%) and uses up to 60% fewer tokens to solve harder problems. The reasoning model MAI-Thinking-1 uses a 35-billion-parameter mixture-of-experts (MoE) architecture, with Microsoft reporting scores of 97.0% on AIME 2025 and 94.5% on AIME 2026.
Mustafa Suleyman, CEO of Microsoft AI, said that after the models were fine-tuned for the needs of consulting firm McKinsey, they outperformed OpenAI's GPT-5.5 while delivering 10 times better cost efficiency. In blind side-by-side tests run by independent human-evaluation partner Surge, MAI-Thinking-1 was preferred over Claude Sonnet 4.6 and matched Claude Opus 4.6 on SWE-Bench Pro coding tasks. Microsoft stressed that running the models on its own Azure cloud infrastructure lets it avoid payments to third parties.
The step comes as the relationship between Microsoft and OpenAI—long its largest backer—increasingly shifts toward competition. As the cost of leading models rises, Microsoft plans to pass on savings to developers by offering its own models. The company introduced a total of seven new MAI models at Build 2026, a move widely seen as concrete evidence of its strategy to become a self-sufficient player in artificial intelligence.