Meta is exiting the same vendor twice, and the two exits carry very different costs. Pulling engineers off Claude Code costs habit: Meta removed the token leaderboards once costs climbed [9], shipped an update to Muse Code, the coding tool it launched in August, and its employees are increasingly reaching for that instead [19]. Pulling Anthropic out of Hatch costs evals. Four people told the Times that the consumer agent was powered and tested on Anthropic's models [13], and one said the version that debuts publicly will run on Meta's own latest model [14]. Every judgement anyone inside Meta formed about how Hatch behaves was formed against a model that will not be in the launch build, on a product the company has considered pricing at up to $199.99 a month [15].
Renting a competitor's model is often framed as a bridge, held until the in-house model lands. Inside Meta, engineers instead competed on internal leaderboards over who could burn the most Anthropic tokens, a practice the industry calls tokenmaxxing [8]. Token burn is this cycle's stand-in for usage depth, and what it measures is the invoice. It does not say whether the pull request landed, or what the same work would have cost on a smaller model. Meta's correction was administrative: in June it told employees it was on track to spend billions on AI this year and would build a better system to manage that spending [10], which is roughly the move Microsoft made in August when it set division-level budgets [11].
The projection is worth dividing out. Up to $10bn a year is about $833m a month [23], against a current rate of hundreds of millions a month after this summer's cutbacks [12]. The internal ceiling was close to where the spending was already heading, annualised. Set against Anthropic's July estimate that its yearly revenue would pass $65bn [2], a single $10bn customer accounts for about 15% of the top line [24].
That scale is what turned the line item into a competitive question, not just a procurement one. Nat Friedman, Meta's head of AI product, has told some employees that using only Meta's own coding tools or OpenAI's could lower Anthropic's revenue ahead of an offering [18] that TNW reports could value the company at $2tn [17]. The traffic runs in both directions, and cross-buying is normal at this altitude: Meta already pays Microsoft hundreds of millions a year to rent models [5], Google and Amazon have committed $73bn to Anthropic while building rivals [6], and in June Anthropic approached Meta with a proposal to buy up to $10bn in computing [22].
For anyone making this call at a smaller scale, the sorting question is swap cost, feature by feature. Put each place you pay a competitor's model on two axes: whether a model change is visible to the user, and whether your read on quality is tied to that particular model. Internal tooling sits in the cheap corner, where an exit costs retraining and a fortnight of grumbling, and it deserves metering from the first week rather than a leaderboard. A shipped feature whose behaviour your testers have internalised sits in the expensive corner, where the exit buys you a new eval set and an unknown quality baseline. Meta worked both corners this year, deleting a leaderboard [9] and rebuilding the model underneath a product it has not shipped [14].
