Microsoft AI Chief Says Company Was 'Set Free' From OpenAI to Chase Superintelligence

Microsoft Declares Independence

Microsoft AI chief Mustafa Suleyman has declared that a renegotiated deal with OpenAI has "set free" the company to independently pursue superintelligence, marking a turning point in one of the most consequential partnerships in technology. Speaking to the Financial Times, Suleyman said Microsoft's superintelligence team is now targeting enterprise use cases, software developers, and coding, an approach that closely mirrors Anthropic's strategy, while being "less concerned" about the consumer-focused tactics of Google, Meta, and OpenAI.

A New Model Family

The declaration came as Microsoft unveiled a family of seven in-house AI models at its Build 2026 developer conference in San Francisco on 2 June. The flagship, MAI-Thinking-1, is a sparse Mixture-of-Experts reasoning model with roughly 1 trillion total parameters but only 35 billion active during inference, paired with a 256,000-token context window. Microsoft says it matches Anthropic's Claude Opus 4.6 on the SWE-Bench Pro coding benchmark and scores 97% on the AIME 2025 maths reasoning test. The suite also includes MAI-Code-1-Flash, tuned for Visual Studio Code and GitHub Copilot, plus models for image generation, transcription, and voice synthesis. Crucially, Microsoft emphasised that every model was trained from scratch on commercially licensed data, with no distillation from third-party systems.

The Restructured Deal

The shift follows a restructured agreement finalised in October 2025, in which OpenAI converted its for-profit arm into a public benefit corporation. Under the revised terms, Microsoft received a 27% ownership stake valued at around $135 billion and retained access to OpenAI's technology through to 2032, including any models that reach artificial general intelligence. Critically, the deal lifted restrictions that had previously stopped Microsoft from training its own frontier models. Suleyman first signalled the new direction in November 2025 when he announced the MAI Superintelligence Team, led by chief scientist Karen Simonyan, to pursue what he calls "humanist superintelligence."

Enterprise Ambitions and Custom Silicon

Microsoft's enterprise focus extends beyond models. The company is building custom silicon, with its Maia 200 accelerator reportedly 30% more cost-efficient than Nvidia's GB200. MAI-Thinking-1 is available in private preview on Microsoft Foundry, with a public preview expected shortly. Suleyman has also predicted that AI will achieve human-level performance on most professional white-collar tasks within 18 months, and claimed Microsoft's models beat OpenAI's GPT-5.5 on McKinsey benchmarks at a tenth of the cost. These moves position Microsoft as both a partner to and a competitor of OpenAI, retaining its stake and model access while building the infrastructure to go it alone.