Meta and Microsoft have reduced their reliance on Anthropic's Claude family of AI models, according to people familiar with the matter. The pullback comes as both companies' internal AI systems have matured to the point where they can handle workloads that previously required an outside model. Neither company has announced the change publicly.
The shift is a blow to Anthropic's enterprise revenue at a time when the company is trying to convince businesses that frontier models are worth renting rather than building. It also underscores a broader tension in the AI industry: the same giants that helped validate third-party model providers are now their most direct competitors.
Why the usage dropped
The reason isn't a falling-out or a pricing dispute. It's simpler than that. Meta and Microsoft have spent the past several years building and refining their own model stacks, and those tools have reached a level of maturity where using Claude for certain tasks no longer makes sense. When an in-house option performs well enough, the calculus shifts: why pay per token for something you can run on your own infrastructure, with your own data, under your own terms?
That logic applies most forcefully at companies with the engineering talent and compute budget to go it alone. Meta and Microsoft are two of the few organizations on the planet that fit that description. For them, the question was never whether they could build a competitive model. It was when the internal version would be good enough. That moment appears to have arrived for at least some of their workloads.
Anthropic's awkward position
Anthropic still has a strong hand. Prediction market traders put the odds of Claude being considered "the best AI model" by November 2026 at 47.5%, which is remarkable given the competition. The company's models remain widely respected for coding, reasoning, and safety work. But respect and revenue don't always move together.
The enterprise AI market is splitting into two camps. One is made up of companies that buy models as a service. The other is made up of companies that build their own. The second camp is small in number but enormous in spending power, and it's the one Meta and Microsoft belong to. Every time a member of that camp trims its external model usage, a meaningful chunk of revenue disappears.
The build-versus-buy calculation
For years, the conventional wisdom in enterprise tech was that buying was faster and cheaper than building. AI has complicated that equation. Training a frontier model costs a fortune, but once the model exists, the marginal cost of using it internally can be far lower than paying an outside vendor. The more a company uses AI, the more the math tilts toward building.
That's the trap Anthropic and its peers face. Their best customers are also their most capable potential competitors. Every contract signed with a tech giant carries an implicit expiration date: the day the giant's internal model gets good enough.
Anthropic hasn't said how much revenue is at stake, and Meta and Microsoft haven't detailed which specific workloads moved in-house. The lack of disclosure makes it hard to size the impact. But the direction is clear.
What to watch
The next signal will come from Anthropic's enterprise customer disclosures, if the company chooses to make them. Watch also for whether other large AI adopters follow Meta and Microsoft in reducing their dependence on outside models. If the trend spreads beyond the handful of firms with world-class AI teams, it becomes a much bigger problem for the entire model-as-a-service business.
For now, Anthropic's best defense is to stay ahead on capability. If Claude remains the model that other models are measured against, customers will keep paying even when they have alternatives. The 47.5% probability on the prediction market says the market thinks that's roughly a coin flip. Anthropic has until November 2026 to prove the odds wrong.
