SATYA NADELLA spent $13.8 billion convincing the world that Microsoft's future was inseparable from OpenAI's. Six years later, his company just shipped a product feature that runs Anthropic's Claude as a peer reviewer of OpenAI's GPT — inside Microsoft's own flagship productivity suite — and called it a better experience than either model working alone. Somewhere in the fine print of the AI revolution, the most expensive partnership in technology history is being quietly renegotiated in public.
The announcement, made today, introduced two multi-model capabilities to Microsoft 365 Copilot's Researcher agent. Critique pairs an OpenAI model for generation with an Anthropic model for evaluation, separating the thinker from the checker; Council runs both models simultaneously and lets a judge synthesize where they agree, diverge, and offer unique insights. Microsoft claims Critique beats the best single-model deep research system by nearly 14% on the DRACO benchmark, a hundred-task evaluation spanning ten domains. The message is unsubtle: no single model — including the one made by the company Redmond has backed more heavily than any technology partner in its history — is sufficient on its own.
This is not an isolated product update. It is the latest move in a methodical pivot toward model pluralism that has accelerated since the OpenAI partnership was renegotiated last October. In November 2025, Microsoft invested up to $5 billion in Anthropic, which in turn committed $30 billion in Azure compute purchases. By January 2026, Microsoft was on track to spend roughly $500 million per year licensing Anthropic's models for its own products. Microsoft's Azure sales force now gets the same quota credit for selling Anthropic models as it does for OpenAI's — a seemingly mundane incentive change that speaks volumes about strategic intent.
Embrace, extend, abstract
Yet the most revealing thing about Microsoft's multi-model turn is how deeply it rhymes with the company's entire competitive history. Microsoft has always been a platform company, not a components company. Windows sat above commodity hardware. Office sat above commodity document formats. Azure sat above commodity servers. In each case, Microsoft's playbook was the same: let others compete to provide the interchangeable inputs, then own the orchestration layer that enterprises cannot live without. The $13.8 billion was not a declaration of fealty to OpenAI; it was the cost of learning the AI stack well enough to commoditize it.
The numbers tell that story clearly. Microsoft holds a 27% stake in OpenAI valued at roughly $135 billion — a formidable return on paper. But the restructured partnership agreement, signed last October, loosened the exclusivity that once bound the two companies together. OpenAI can now serve non-API products on any cloud provider. Microsoft can independently pursue artificial general intelligence alone or with third parties. OpenAI committed to purchasing $250 billion of Azure services, but Microsoft surrendered its right of first refusal as a compute provider. Read closely, these terms describe not a marriage but a lucrative commercial arrangement with built-in optionality — precisely the kind of deal a platform company writes when it expects the supplier market to fragment.
The operational evidence is even more telling. Copilot Studio now offers enterprises a model picker with OpenAI, Anthropic, and bring-your-own-model options from a catalog of over 1,900 models. GitHub Copilot supports GPT, Claude, and Gemini variants across its tiers. Microsoft's own in-house Phi models handle edge and on-device tasks. As one Microsoft executive put it with surprising candor, "Every 60 days at least, there's a new king of the hill." That is not something you say about a partner you expect to remain permanently ahead.
Still, there is a counterargument worth pressure-testing. Model-agnosticism works beautifully as a cloud sales pitch — enterprises love optionality — but it introduces real complexity at the product level. Critique's architecture concedes that frontier models still cannot reliably judge their own output, so Microsoft had to build an entire second-model layer to enforce accuracy. That is an elegant engineering solution, but it is also an admission: if the AI industry's autonomous-agent narrative were playing out on schedule, you would not need one model to babysit another. The gap between what AI can produce and what enterprises will trust remains wide enough to require an entire product category — evaluation-as-a-service — to bridge it.
The question now is whether the orchestration layer becomes as lucrative for Microsoft as the model layer has been for OpenAI and Anthropic, or whether it becomes a thin margin pass-through. With only 15 million paid Copilot seats — a mere 3.3% of Microsoft's 450-million-strong commercial installed base — the enterprise AI revolution remains stubbornly nascent. Multi-model sophistication is a compelling pitch for the CIO already buying; it does little to convert the 96.7% who have not yet seen the point. Nadella's $13 billion bought Microsoft the pole position in a race where most of the runners have not yet left the starting blocks. Whether they eventually do may matter more than which model carries them across the finish line. ■
For more, join 75,000 subscribers getting tech's favorite brief here
