Sources: Microsoft is working on adding more first- and third-party AI models to power 365 Copilot, to lower its reliance on OpenAI amid cost and speed concerns
Sam Altman and the gang absolutely stiffed Microsoft so bad. How have they basically part funded OpenAi and yet been given such a weak version of the model to embed inside 365?! [embedded post] X: Andrew Curran / @andrewcurran_ : Since last November, this was always only a matter of time. Interestingly, the report says they are looking at both ‘internal and third-party’ models. [image] Barry Schwartz / @rustybrick : ooo Microsoft and OpenAI drama?
Context & Ripple Effects
Microsoft had already been reported to be pursuing cheaper, more efficient models for AI features, including systems designed to mimic more capable OpenAI models. This report extends that cost-control logic from infrastructure into the model choices behind a major productivity product.
The later coverage arc makes the direction clearer: Microsoft was reported to be testing its MAI model family in Copilot and, later still, using Anthropic for selected Office capabilities after task-specific performance comparisons. This initial move matters as the opening of a more modular model strategy rather than a one-off vendor dispute.
First-order effects
- Microsoft can evaluate first-party and outside models for 365 Copilot workloads where OpenAI models are too costly or slow, potentially changing which model powers a given feature.
- OpenAI faces a less exclusive role in a high-profile Microsoft product, while Microsoft gains more leverage over performance, availability, and inference costs.
Second-order effects
- Other model providers gain a route into enterprise productivity software if they can demonstrate better cost, speed, or task performance for specific Copilot functions.
- A multi-model Copilot architecture raises the importance of routing, evaluation, and integration capabilities: Microsoft must select models by workload rather than treat model access as a single procurement decision.
Third-order effects
- If this pattern persists, enterprise AI products will increasingly compete on the ability to combine and manage multiple models, not solely on affiliation with one frontier-model supplier.
- The shift could rebalance bargaining power toward large application distributors that can switch among providers or deploy their own models, though the extent depends on whether alternatives reliably meet product-quality requirements.
The trend: This is one data point in the unbundling of enterprise AI stacks, as major software platforms diversify model supply to improve economics and reduce dependence on a single lab.