Microsoft Says It Was 'Set Free' From OpenAI Six Months Ago — Now It's Building Its Own Frontier Models

A contractual renegotiation quietly removed restrictions on Microsoft's in-house AI research. The seven-model MAI family announced at Build 2026 is the first tangible result.

Written by Lena Armitage · Bureau Tech · Jun 12, 2026

The contract clause nobody talked about

For years, Microsoft's AI strategy was, functionally, OpenAI's AI strategy. The partnership — built on a cumulative investment exceeding $13 billion — gave Microsoft early access to frontier models and made it the exclusive cloud provider for OpenAI's infrastructure. What was less visible: the agreement also restricted what Microsoft could build on its own, including a cap on the compute threshold, measured in FLOPS, for any model it trained internally.

That restriction is now gone. Mustafa Suleyman, CEO of Microsoft AI, told VentureBeat at Build 2026 that a contract revision roughly six months ago formally cleared the way for his division to pursue what he calls "superintelligence" using Microsoft's own researchers, data pipelines, and silicon. "We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence," he said. "So this is very early days."

The revised terms were first reported by Fortune and Axios in November. Suleyman's Build comments are the most direct public acknowledgment of what those changes actually permitted.

Seven models, one signal

The clearest evidence of the shift is the MAI model family, announced at Build 2026. Seven models, developed entirely in-house by Microsoft's AI Superintelligence Team, cover reasoning, code generation, image creation, transcription, and voice synthesis.

The flagship, MAI-Thinking-1, is a 35-billion-active-parameter reasoning model. Microsoft says it matches leading models in its weight class on software engineering benchmarks and shows strong mathematical reasoning. One claim Suleyman repeated: the model was trained from scratch on commercially licensed data, without distillation from other labs' outputs — a practice that has become common in the industry and is increasingly contested legally.

The benchmark comparisons Microsoft offered are worth reading carefully. "Matches leading models in its weight class" is a relative claim bounded by weight class, not an assertion of overall frontier performance. The company has not published independent third-party evaluations.

The rest of the family is designed for enterprise deployment: MAI-Code-1-Flash for GitHub Copilot and VS Code; MAI-Image-2.5 for text-to-image and image editing; MAI-Transcribe-1.5, which Microsoft claims is the most accurate transcription model available across 43 languages; and MAI-Voice-2 for multilingual speech generation. All ship through Microsoft Foundry, the company's model-hosting infrastructure.

The enterprise tuning play

Alongside the models, Microsoft announced Frontier Tuning — a system that lets enterprise customers adapt MAI models to their own workflows, terminology, and data, inside their own compliance boundaries. The mechanism uses reinforcement learning environments Microsoft describes as "training gyms" that let agents learn from real workplace tasks without touching production systems.

Microsoft's own numbers: an MAI model tuned for Excel reportedly matches GPT performance at up to ten times greater efficiency. One unnamed enterprise customer achieved the highest win rate of any model tested at roughly one-tenth the cost. These figures come from Microsoft and have not been independently verified.

Early Frontier Tuning partners include Mayo Clinic, which is co-developing a healthcare-specific model on de-identified clinical data; EY, tuning a tax-advisory agent for 75,000 professionals; and Pearson, using tuned models for learning-science feedback.

The hardware argument

Suleyman was unusually specific about compute economics. Microsoft's second-generation custom chip, Maia 200, is already running in production in Iowa and Arizona data centers. He claims it is 30 percent more cost-efficient than Nvidia's GB200, and that co-optimizing MAI models to run natively on Maia silicon adds another 1.4x improvement in performance per watt.

Microsoft simultaneously describes itself as the world's largest buyer of Nvidia GB200s and GB300s. The two positions are not contradictory — custom silicon typically handles specific workloads while GPU clusters handle breadth — but the cost-efficiency claims for Maia 200 are Microsoft's own and have not been independently benchmarked.

What 'self-sufficiency' actually means by 2030

Suleyman is not claiming Microsoft has already built a frontier lab. He is arguing it is building one. "Our job is to make sure that when we look out to 2030 and beyond, we have the capacity not just to buy models from third parties, but to build the absolute frontier, the best models in the world," he said. "That's a long transition."

The OpenAI partnership continues. Copilot and Azure AI services still run on OpenAI models. Suleyman described Microsoft's current model portfolio — OpenAI, Anthropic, and thousands of models in Foundry — as optionality, not a gap to fill urgently.

But the direction is clear. Microsoft is constructing a vertically integrated stack: its own models, its own chips, its own reinforcement learning infrastructure, its own enterprise tuning layer. Whether that stack produces frontier-competitive models by 2030 depends on execution that no announcement can guarantee.

Key takeaways

A renegotiated contract with OpenAI — reported by Fortune and Axios in November — removed a cap on how large a model Microsoft could train, measured in FLOPS, and formally permitted the company to pursue its own superintelligence research.
Microsoft's MAI family includes seven models spanning reasoning, code, image generation, transcription, and voice; the flagship MAI-Thinking-1 is a 35-billion-active-parameter reasoning model Microsoft says was trained from scratch on commercially licensed data without distillation from third-party models.
Suleyman claims Microsoft's second-generation custom chip, Maia 200, is 30 percent more cost-efficient than Nvidia's GB200 — a figure that, if it holds at scale, would give Microsoft a structural cost advantage in running its own models.
A new capability called Frontier Tuning lets enterprise customers customize MAI models using proprietary data inside their own compliance boundaries; early partners include Mayo Clinic, EY, and Pearson.
Suleyman argues AI models are not commoditizing, staking the case on 'quality tokens' — the curation, licensing, and composition of training data — rather than raw compute scale.

FAQ

What did the renegotiated Microsoft-OpenAI contract actually change?

According to reporting by Fortune and Axios in November, and confirmed by Suleyman's comments at Build 2026, the revised agreement removed restrictions that had barred Microsoft from pursuing its own AGI research and capped the compute scale — measured in FLOPS — of any model Microsoft could train internally. Microsoft can now build frontier-scale models using its own researchers and infrastructure.

Is Microsoft replacing OpenAI as its primary AI provider?

Not immediately. Suleyman described the current relationship as one of 'abundance' — Microsoft still relies on OpenAI models for Copilot and Azure AI services, and also works with Anthropic and thousands of other models through Foundry. The stated goal is self-sufficiency at the frontier by 2030, not a near-term break.

What is MAI-Thinking-1 and how does it compare to other models?

MAI-Thinking-1 is a 35-billion-active-parameter reasoning model — 'active parameters' refers to the subset of a model's parameters engaged during any given inference pass, a common architecture in mixture-of-experts models. Microsoft says it matches leading models in its weight class on software engineering benchmarks. The company has not released independent third-party benchmark results, so direct comparisons to models outside its weight class are not yet supported by the available data.

What is Frontier Tuning and who is it for?

Frontier Tuning is a Microsoft system that lets enterprise customers customize MAI models using their own proprietary data and workflows, inside their own compliance boundaries. It uses reinforcement learning environments — Microsoft calls them 'training gyms' — so agents can learn from real tasks without affecting live production systems. Early partners include Mayo Clinic, EY, Land O'Lakes, and Pearson.

What is Maia 200 and why does it matter?

Maia 200 is Microsoft's second-generation custom AI accelerator chip, already running in production data centers in Iowa and Arizona. Suleyman claims it is 30 percent more cost-efficient than Nvidia's GB200 and delivers an additional 1.4x performance-per-watt improvement when running MAI models natively. These figures are Microsoft's own; independent benchmarks have not been published.

Citations

Microsoft AI chief says company was 'set free' from OpenAI to pursue superintelligenceSuleyman disclosed that a contractual change with OpenAI roughly six months ago granted Microsoft AI the formal authority to pursue superintelligence using its own researchers, data pipelines, and custom silicon.
Microsoft and OpenAI revise partnership agreement (Fortune / Axios reporting, November 2025)Fortune and Axios reported in November that a revised deal with OpenAI removed restrictions on Microsoft's ability to train large-scale models and pursue its own AGI research.
Microsoft Build 2026 MAI model family announcementMicrosoft announced seven in-house MAI models at Build 2026, including MAI-Thinking-1, a 35-billion-active-parameter reasoning model trained from scratch on commercially licensed data.
Mustafa Suleyman blog post accompanying Build 2026 announcementsSuleyman wrote that Microsoft trains its reasoning models from scratch, does not distill from other labs, and does not rely on unlicensed or opaque data.