Microsoft's MAI-Thinking-1 aims to ease legal fears with clean training data

At its Build conference, Microsoft dropped a family of seven new AI models, headlined by the 35-billion-parameter MAI-Thinking-1 — the company's first reasoning model. AI lead Mustafa Suleyman claimed it outperforms Anthropic's Claude Sonnet 4.6 in overall quality, scoring 97% on the AIME math benchmark and 53% on SWE-Bench Pro for complex coding (vs. Claude Opus 4.6 at 51.9%, but behind OpenAI's GPT-5.4 at 59.1%). Critically, MAI-Image-1 was trained entirely from scratch with zero distillation — meaning it didn't use other models to generate training data — and with commercially licensed data, giving enterprises a clear legal lineage. This addresses a growing fear among businesses: using models trained with questionable data sources could lead to copyright lawsuits.

Alongside the reasoning model, Microsoft introduced MAI-Image-2.5 (which ranks #3 on Arena.AI text-to-image), MAI-Transcribe-1.5 (claimed as the best transcription model), MAI-Voice-2 and Mai-Voice-2 Flash, and MAI-Code-1-Flash. The announcement marks the biggest in-house AI push from Microsoft since the August launch of MAI-Voice-1, signaling a growing distance from its OpenAI partnership. The company brands its AI vision as "humanist superintelligence," aiming to keep humans in control while tackling global challenges — and now, with a focus on data provenance, it's giving enterprise customers a reason to trust its models over rivals'.