Mustafa Suleyman says Microsoft was “set free” from OpenAI six months ago

For three years, Microsoft’s AI story has been inseparable from OpenAI. The partnership, cemented by a cumulative investment exceeding $13 billion, gave Microsoft early access to advanced models, accelerating Copilot’s march into the enterprise and adding hundreds of billions to Microsoft’s market capitalization. Now, Mustafa Suleyman says that dependency has been formally loosened.

In an exclusive sit-down interview with VentureBeat at Microsoft Build 2026, Suleyman said his division was “set free” from its contract with OpenAI roughly six months ago, granting formal authority to pursue “superintelligence.” He didn’t frame it as a dramatic break with OpenAI. He called it early days, and more importantly, a shift in what Microsoft can build with its own researchers, its own data pipelines, and its own custom silicon. In practical terms: Microsoft is using the same enterprise urgency that made Copilot mainstream to build toward a future where the foundational model layer is not just sourced, but also manufactured.

The most visible evidence of that shift arrived the same day. Microsoft announced a family of seven new AI models developed entirely in-house by its AI Superintelligence Team. Branded under the “MAI” family name, the releases cover reasoning, code generation, image creation, transcription, and voice synthesis. The flagship, MAI-Thinking-1, is described as a 35-billion-active-parameter reasoning model. Microsoft says it matches leading models in its weight class on key software engineering benchmarks and demonstrates advanced mathematical reasoning.

What makes MAI-Thinking-1 strategically interesting is not just its scope, it is the training recipe Suleyman emphasized. In a blog post, he wrote that the model was trained from scratch on clean, commercially licensed data, and that Microsoft “doesn’t distill from other labs and… doesn’t rely on unlicensed or opaque data.” That is a direct contrast, at least implicitly, to the broader industry practice of distilling from competitor frontier model outputs to achieve cheaper training paths. If you are an enterprise buyer, the subtext is simple: your risk posture around data provenance and model provenance may matter as much as benchmarks. And if you are a board member, it is a reminder that platform-level control increasingly includes data and training governance, not just distribution.

The rest of the MAI portfolio is built for multimodal enterprise deployment. Microsoft says MAI-Code-1-Flash is a lightweight coding model designed specifically for GitHub Copilot and VS Code. MAI-Image-2.5 supports text-to-image and image editing. MAI-Transcribe-1.5 claims to be the most accurate transcription model available and operates across 43 languages. MAI-Voice-2 is a multilingual speech-generation system. These models ship through Microsoft Foundry, the company’s model-hosting and deployment infrastructure. For the first time, developers can tune model weights themselves through third-party platforms including OpenRouter, Fireworks, and Baseten. Even with all that, Suleyman made clear that these seven models are a proof of concept. The real transition is institutional: building the “capacity… to build the absolute frontier” when Microsoft looks out to 2030 and beyond, not just purchasing models from others.

So what does “set free” actually mean in contract terms? The arrangement Microsoft’s partnership with OpenAI has long embodied an unusual architecture. When Microsoft invested billions into OpenAI beginning in 2019, OpenAI would build frontier models, while Microsoft served as the exclusive cloud provider, integrating and reselling them through Azure. That gave Microsoft extraordinary commercial leverage, but it also created dependency. The original deal explicitly barred Microsoft from pursuing its own AGI research and capped how large a model it could train, restricting systems beyond a computing threshold measured in FLOPS.

Those restrictions, VentureBeat reports via Fortune and Axios reporting in November, were removed in a renegotiated deal with OpenAI. That is the regulatory-adjacent part executives should pay attention to: this is governance through contract, not regulation by statute. But the effect is similar. The contract defines what “frontier-scale” experimentation is allowed, when, and by whom. In Suleyman’s telling from that same context, the revised “best-of-both environment” lets Microsoft pursue its own superintelligence while still working closely with OpenAI. In the interview, he reinforced a posture of abundance rather than scarcity: there’s no urgent need to fill gaps in three or six months because Microsoft has OpenAI, Anthropic, and thousands of models inside Foundry.

That brings us to the commercial logic Microsoft is now pushing: the shift from model access to enterprise control. Suleyman introduced Frontier Tuning, announced alongside the models at Build. Frontier Tuning allows enterprise customers to customize MAI models using their own proprietary data, workflows, and domain terminology within their own secure compliance boundary. It uses reinforcement learning environments, which Microsoft calls “training gyms for AI,” where agents learn from real workplace tasks without affecting production systems. Microsoft claims an Excel-tuned MAI model matches GPT 5.4 performance while operating at up to ten times greater efficiency. For early enterprise adopters, Microsoft says tuned models achieved the highest win rate of any model tested at roughly one-tenth the cost, in an example for an unnamed organization.

Suleyman connected that directly to a strategic pivot in what AI is supposed to do. “We’ve basically moved beyond just conversation,” he told VentureBeat. “Now we’re moving to action.” He laid out a progression from IQ (factual intelligence) to EQ (emotional intelligence, following tone and style instructions) to AQ, the “Actions Quotient.” The future AI agents described here are not just chatbots with better answers. They are positioned to log into enterprise software, navigate multi-application workflows, and execute tasks across tools like Excel, Word, Teams, Jira, and Adobe InDesign.

This is why the “set free” moment matters beyond one executive headline. If Microsoft can move from dependence on partner frontier models to in-house frontier building, then its enterprise roadmap becomes less vulnerable to contract constraints and more controllable end to end: data, training, deployment, and customization. For other leaders in tech and the boards that oversee spend and risk, the second-order implication is clear. The winning platform will increasingly be the one that can deliver not only intelligence, but reliable action under enterprise governance, without waiting on someone else’s upstream releases.