The one-model bet just died

Ninety-nine dollars a month. That's what Microsoft is charging enterprise customers for its new 365 E7 Frontier Suite, which bundles Copilot, Agent 365 governance, and enterprise security into a single package. Buried in the product spec is the detail that matters: Copilot Wave 3 now routes tasks between Claude and GPT models automatically, choosing whichever performs best for each job. Microsoft's most important AI product just declared that no single model deserves exclusive access to the enterprise.

This would have been heresy eighteen months ago. Microsoft had invested $13 billion in OpenAI. The partnership was supposed to be the defining alliance of the AI era. Now Microsoft is openly routing work to Anthropic's Claude when it outperforms GPT on a given task, inside the same workflow, at the same price point. The message to OpenAI is diplomatic but unmistakable: you're a supplier, not a partner.

The same instinct showed up in a very different context. CNN reported that the Pentagon signed classified AI deals with seven companies (OpenAI, Google, Microsoft, Amazon, Nvidia, SpaceX, and Reflection AI), clearing them all for Impact Level 6 and 7 networks. The Pentagon's CTO said explicitly that relying on a single provider would have been "irresponsible." Defence procurement has always moved toward multi-vendor strategies eventually, but this happened fast. The Department of Defense reached the same architectural conclusion as Microsoft's product team in roughly the same week: model monogamy is a liability.

(The Pentagon story has a darker footnote. Anthropic was frozen out entirely, designated a supply-chain risk, after refusing to permit unrestricted use of Claude in autonomous weapons systems and mass surveillance. What it means when principled refusal gets you blacklisted rather than respected deserves its own editorial, but the multi-vendor point stands regardless.)

The price signal

Then there's what happened on the supply side. VentureBeat reported that xAI completed its Grok 4.3 API rollout with input prices slashed roughly 40%, a context window expanded to one million tokens, and native video input added — all released silently, with no press event. When a company cuts prices by 40% and doesn't even bother to announce it, the competitive dynamics have shifted from differentiation to commoditisation. xAI isn't trying to convince you Grok is the best model. It's trying to make sure it stays on the shortlist.

This is what commoditisation looks like in practice. In the steel industry, the shift from proprietary alloys to standardised grades in the mid-twentieth century didn't kill steelmakers. It moved the value from the material to the fabrication: the companies that shaped and assembled the steel, not the ones that poured it. The AI model layer is going through the same transition, except compressed into quarters instead of decades. Microsoft is the fabricator. The labs are the mills.

The practical implication for anyone building products on AI: if you've designed your architecture around a single model provider, you've built a dependency that your largest potential customers are actively trying to avoid. Microsoft's enterprise buyers are paying $99 a month specifically so they don't have to pick one model. The Pentagon is signing seven contracts specifically so it doesn't depend on one vendor. Your customers will expect the same optionality.

Design for portability. The one-model era didn't decline gradually. It ended in a week.

---