Microsoft's seven new MAI models make a lot more sense once you read the OpenAI contract behind them

Seven MAI models in one keynote, and almost no context

On June 2, 2026, at Microsoft's Build conference, Microsoft put seven new in-house AI models in front of developers in a single announcement block: MAI-Thinking-1 for reasoning, MAI-Code-1-Flash for coding, two image models, a 43-language transcription model, and two voice models. Most of the coverage that followed read like a spec sheet: parameter counts, leaderboard placements, a claimed parity with Claude Opus 4.6 on SWE-Bench Pro for the flagship reasoning model. Almost none of it asked the more interesting question. Why would Microsoft, a company whose AI products have run almost entirely on OpenAI's models since 2019, suddenly ship seven MAI models of its own, led by a reasoning model Microsoft says was trained without OpenAI data and without distillation from GPT?

The answer isn't in the keynote. It's in a contract amendment Microsoft and OpenAI signed five weeks earlier, and almost nobody read the two events side by side.

The eight months nobody connected

Lay out the last eight months in order and the MAI launch stops looking like a roadmap milestone and starts looking like a hedge.

• Oct 28, 2025: OpenAI completed its for-profit recapitalization. Microsoft converted its investment into roughly 27% of the new OpenAI Group PBC, a stake valued near $135 billion. In return, OpenAI committed to an incremental $250 billion in Azure spending, and Microsoft gave up its right of first refusal on OpenAI’s compute purchases, freeing OpenAI to buy GPU capacity from any cloud provider.
• Nov 3, 2025: OpenAI signed a separate $38 billion, multi-year compute deal with AWS, its first major cloud contract with a provider other than Microsoft.
• Sept 2025: OpenAI separately signed a five-year, roughly $300 billion cloud computing agreement with Oracle to expand its Stargate data-center buildout, deepening a compute strategy that no longer runs through one vendor.
• Apr 27, 2026: Microsoft and OpenAI amended the revenue-share terms from the October deal. OpenAI’s payments to Microsoft, previously an open-ended 20% of revenue running until OpenAI’s board declared artificial general intelligence had been reached and an independent expert panel verified it, became a flat $38 billion cap running through 2030, regardless of what OpenAI’s models can do by then. Microsoft has said the cap saves OpenAI roughly $97 billion against the uncapped path. OpenAI also gained the right to serve customers from any cloud provider, not just Azure.
• Jun 2, 2026: Microsoft shipped seven MAI models at Build, led by a reasoning model trained without OpenAI’s technology.

What the April amendment actually changed

The headline most outlets ran was some version of "Microsoft and OpenAI end exclusivity." That is true, but it undersells the more consequential change, which is what happened to the revenue-share clause itself. Under the October 2025 terms, Microsoft’s 20% cut of OpenAI’s revenue was tied to an AGI clause: payments were open-ended until OpenAI’s board declared artificial general intelligence had been reached, a determination that itself required sign-off from an independent expert panel. That clause meant Microsoft’s largest single revenue line tied to OpenAI had no fixed end date, and was, in principle, hostage to a definition neither company fully controlled on its own.

The April amendment replaced that with a number and a date: $38 billion, paid out through 2030, full stop. Converting an open-ended, AGI-contingent revenue stream into a fixed-dollar, fixed-date one matters more than the exclusivity language that made the headlines. It tells both companies exactly how much they owe and are owed, which is the kind of certainty that makes it possible to plan a five-year roadmap, or fund an in-house model-training program, without waiting on a definition nobody has actually agreed on yet.

MAI models aren’t trying to win a benchmark war

Read against that timeline, the seven MAI models look less like a bid to out-perform GPT-5.5 or Gemini 3.5 Pro and more like an exercise in optionality. MAI-Thinking-1, the flagship reasoning model, is a comparatively small 35-billion-parameter model with a 256K context window. Microsoft has published its own benchmark comparison claiming parity with Claude Opus 4.6 on SWE-Bench Pro, a real coding benchmark, but that is a vendor-reported number, not an independent evaluation, and it deserves the same caution any single-vendor benchmark claim does. MAI-Code-1-Flash, a smaller 5-billion-parameter coding model, is already live inside GitHub Copilot and VS Code. The remaining five models cover image generation, transcription across 43 languages, and voice. Notably absent from the lineup: a video generation model, the one category where OpenAI, Google, and several smaller labs have all shipped flagship products this year.

What’s missing is as telling as what shipped. Microsoft has not published API pricing for any of the seven models. The company is positioning them as available not just through Azure AI Foundry but through third-party model routers like OpenRouter, Fireworks, and Baseten, the same distribution channels independent model labs use to reach developers who aren’t already committed to a single cloud. That is not the move of a company trying to lock customers deeper into Azure. It is the move of a company making sure it has a credible model story even if its relationship with OpenAI keeps loosening.

The dependency didn’t disappear. It moved up a layer.

None of this means Microsoft and OpenAI are splitting up. Microsoft still owns roughly 27% of OpenAI Group PBC. OpenAI still owes Microsoft $38 billion through 2030 and is still on the hook for $250 billion of incremental Azure spend. Azure is still, by Microsoft’s own description, OpenAI’s priority cloud, meaning new OpenAI products ship there first unless Azure genuinely can’t support them yet. What changed is that neither company’s near-term success is contractually tied to the other’s anymore. OpenAI can scale on AWS or Oracle’s Stargate infrastructure without breaching anything. Microsoft can ship a reasoning model that doesn’t touch OpenAI’s technology without breaching anything either. The companies remain financially intertwined. They are no longer financially dependent.

That distinction is the actual story. Lock-in did not disappear, it moved from the infrastructure layer, who runs the GPUs, up to the ecosystem layer: GitHub Copilot, Microsoft 365 Copilot, the Windows Copilot Runtime, all of which now route between OpenAI and Microsoft’s own models depending on the task. A developer using GitHub Copilot today cannot easily tell, without reading the documentation closely, which requests go to GPT-5.5 and which go to MAI-Code-1-Flash. The vendor relationship moved out of the contract and into the product surface, where it is harder to see and harder to renegotiate around.

What this changes if you’re building on Azure right now

For a team running production AI features on Azure today, nothing breaks. Azure OpenAI Service still runs OpenAI’s actual models, on the same SLAs as before. Three things are worth re-evaluating at the next architecture review, though:

• Azure AI Foundry is now a genuine multi-model marketplace, not an OpenAI proxy with extra steps. MAI models sit next to OpenAI’s in the same catalogue, and once pricing publishes, the cost comparison may not favor whichever model a team defaulted to a year ago.
• The two roadmaps will diverge on independent schedules. A capability that ships on OpenAI’s own platform first may now take longer to reach Azure than it used to, since "priority" cloud status is a weaker commitment than the old exclusivity arrangement. Checking release notes rather than assuming day-one parity is worth the five minutes.
• The vendor-risk shape changed. A year ago, betting on Azure OpenAI Service was effectively one bet on a combined Microsoft-OpenAI entity. It is now two bets on two companies that are financially independent of each other below a fixed dollar ceiling. That is not necessarily worse, but it is a different risk to model, and most vendor-risk reviews have not caught up to it yet.

Both companies are hedging. That’s the actual headline.

OpenAI’s diversification away from Microsoft, the AWS deal, the Oracle Stargate commitment, the agreements with Google Cloud and CoreWeave, gets covered as OpenAI outgrowing its first investor. Microsoft’s MAI launch gets covered as Microsoft building competitive AI products. Both framings are true, narrowly. But put the contract dates next to the product dates and what they actually describe is two companies that spent 2023 through 2025 betting heavily on each other’s exclusivity, and spent the first half of 2026 quietly pricing in what happens if that exclusivity isn’t there anymore. The AI infrastructure market’s most important partnership did not end in April. It just stopped assuming it had to last forever.