Key points
- Internal projections showed Meta (META) might spend up to $10 billion annually on Anthropic's tools, according to the New York Times.
- Meta is already spending hundreds of millions each month and relied on Anthropic's models while developing and testing its Hatch agent.
- In June, Meta restricted its own engineers from Claude Code, and Zuckerberg criticized Anthropic and OpenAI in an essay on August 10.
Meta (META) spent this summer making the public case against Anthropic, and it spent the same months projecting internally that it could pay Anthropic as much as $10 billion a year. The New York Times reported Thursday that Meta has become one of Anthropic's largest customers and is still spending hundreds of millions of dollars a month on its tools.
According to the Times, Meta used Anthropic's models to help power and test Hatch, the personal AI agent it plans to launch in the coming weeks, while its own teams were building products that compete with Anthropic directly. Employees have been shifting over to Muse Code, the in-house coding agent Meta shipped in August alongside Muse Spark 1.2. The Information reported earlier this month that Meta was targeting October for its next frontier model, known internally as Watermelon, and the Times said Thursday that it has been pushed to October at the earliest.
The $10 billion is a projection out of Meta's own planning documents, not a signed contract. It is also a company-wide number for an input Meta buys, not revenue Anthropic has booked.
What Meta has said about Anthropic
On August 10, Mark Zuckerberg published a 6,500-word essay arguing that closed model developers such as Anthropic and OpenAI are centralizing power in a way that holds back innovation and creates risk for the industry, while Meta releases Llama and the Muse line as open weights.
In late June, reports said engineers on Meta's Applied AI team could no longer use Claude Code or OpenAI's Codex without approval, because Meta's internal systems interact with those models aggressively enough to raise a distillation risk. Distillation is training a smaller model to copy the behavior of a larger one, and doing it to a competitor's model without permission raises legal and ethical concerns. Meta has not publicly confirmed the policy, and it was never clear whether it reached beyond Applied AI.
An internal leaderboard already showed the scale in April
A Meta employee built a dashboard called Claudeonomics, named after Anthropic's product, that ranked the company's heaviest AI users with titles like "Token Legend." It covered more than 85,000 employees and showed 60.2 trillion tokens consumed in a 30-day window, with the top individual averaging 281 billion tokens.
At Anthropic's list API prices that volume runs near $900 million, though The Pragmatic Engineer noted Meta likely buys at a volume discount and put the real figure at $100 million or more a month. The dashboard came down two days after the numbers were written up. "It was meant to be a fun way for people to look at tokens, but due to data from this dashboard being shared externally, we've made the decision to shutter Claudeonomics for now," the employee who built it wrote in the shutdown message, according to Fortune. Neither Zuckerberg nor CTO Andrew Bosworth ranked in the top 250.
Meta also wants to sell Anthropic computing power
In July, Meta entered early talks to lease computing power from its data centers to Anthropic in a deal the Times valued at up to $10 billion over two years. CNBC confirmed the discussions were preliminary and both companies declined to comment. Nothing has been announced since.
Anthropic already buys compute from Google (GOOGL), Amazon (AMZN), Microsoft (MSFT) and SpaceX, and it filed a confidential draft S-1 with the SEC on June 1 that it has not yet followed with a public filing, while investors have discussed a valuation near $2 trillion.
META was trading 0.75% lower at $571.82 at 11:12 a.m. Eastern Thursday than the previous close of $576.14 on Wednesday.
Meta has guided to capital spending of as much as $145 billion this year, and it took a $10 billion charge into the third quarter for the teen-safety settlement it agreed to on August 26. Neither figure includes what it pays Anthropic.
Meta has also been hunting for new places to raise the money. Meta made a credit rating inquiry in South Korea this week about a won-denominated bond sale to fund AI spending, which Korean media put near 1 trillion won, or $720 million.
