Microsoft Showed Up to Build With Seven of Its Own Models. The OpenAI Era Is Over.

The models that matter

Spec sheets from keynotes deserve skepticism, but two of these have claims you can check today.

MAI-Code-1-Flash is live in the GitHub Copilot model picker as of Monday. Microsoft's pitch is efficiency: comparable output at 60 percent fewer tokens than competing models. After last week's Copilot pricing change, token efficiency stopped being an abstract benchmark and became your bill. A model that does the same work for 60 percent fewer tokens is a 60 percent discount in the new billing world. I ran it on a few refactoring tasks this week; output quality felt a notch below the top of the picker, but the credit burn was visibly lower. For routine work, that trade is going to win a lot of defaults.

MAI-Image-2.5 ranks above Gemini's image model on the Arena leaderboard, which is the first time a Microsoft-trained model has topped a Google one on a public board that people actually watch. MAI-Transcribe-1.5 claims 43 languages at five times the speed of rivals.

The rest of the keynote was the now-standard agent buffet: Project Solara for agent-powered devices, Agent 365 for managing fleets of work agents, Work IQ APIs, and the Majorana 2 quantum chip for the long-horizon crowd.

The contract that explains the strategy

The announcement that actually reframes everything is the one with no demo: a US Department of Defense AI deal covering 2.1 million service members and 770,000 civilian staff, the largest government contract in Microsoft's history.

You cannot run national defense workloads on a model you rent from a partner you're also competing with. Sovereignty over the full stack, silicon to model to deployment, is the price of admission for contracts at this scale. Seen through that lens, the MAI lineup isn't a vanity project or a negotiating tactic. It's a procurement requirement.

What happens to the OpenAI relationship

Nothing dramatic, which is what makes it interesting. Azure still serves OpenAI models. Copilot still offers them in the picker, now next to MAI options with different price-performance profiles. Microsoft remains an investor, and even showed up in Anthropic's Series H last month.

The partnership is becoming what these things always become at scale: a vendor relationship among several. Microsoft's play is the marketplace and the enterprise plumbing, not allegiance to any one model family, including its own.

The model picker is the tell

Microsoft doesn't need MAI models to beat GPT-5 or Claude at the frontier. It needs them to be good enough and cheap enough that defaults inside Copilot, Windows, and government contracts flow to the house brand while the premium options remain a click away.

Watch where MAI-Code-1-Flash sits in the picker by August: if it's the default for new users, the strategy is working. Good enough, at massive distribution, with 60 percent fewer tokens, is how the next phase of this market gets won.