Meta’s $10 Billion Anthropic Bill Comes Due After the Raid

Meta internally projected it could spend as much as $10 billion a year on Anthropic’s AI models, people familiar with the companies said. Mark Zuckerberg, Meta’s chief executive, had just published a 6,500-word essay that attacked unnamed labs for talk filled with doom, without naming Anthropic or Dario Amodei, its chief executive.

That ceiling showed up after Meta failed to hire Anthropic’s researchers last summer and after an in-house model slipped in training. The company has started moving work onto its own tools, and it is still sending hundreds of millions of dollars a month to the same lab.

Zuckerberg Attacked the Labs His Company Still Pays

On August 10, Zuckerberg argued that most other labs build AI for companies and governments, so the balance of power will favor larger institutions if those labs lead. He wrote that the talk from many people building AI is filled with doom. He did not name Anthropic.

Meta is the company primarily focused on building personal superintelligence for everyone. Most other labs are focused on building AI for companies, governments, or other institutions, so if those labs lead, then the balance of power will favor larger institutions over individuals.

Mark Zuckerberg, chief executive, Meta newsroom essay

The same essay sketches the product Meta calls Hatch: a personal agent that works around the clock on health, money, relationships, and the rest of a user’s life. People familiar with the work said Hatch was powered and tested with Anthropic’s models. The version meant for the public, they said, is supposed to run on Meta’s latest model instead.

Meta and Anthropic declined to comment.

Nine-Figure Offers Barely Moved Anthropic Staff

Last summer Meta tried to hire AI researchers with packages in the nine figures, a push Zuckerberg ran himself by text, email, and dinners in Tahoe and Palo Alto. The company landed names from OpenAI and Google and installed Alexandr Wang, Scale AI’s founder, as chief AI officer after taking a 49 percent stake in Scale for about $14.3 billion. Anthropic was the miss.

THE RAID’S UNEVEN HAUL

• OpenAI and Google: Meta hired high-profile researchers from both, including people who had worked on Gemini and on OpenAI’s recent model stack.
• Anthropic: Three people familiar with the companies said Meta largely failed to recruit there, even with cash-heavy offers.
• The rare cross: Joel Pobar, who had worked on inference at Anthropic after 11 years at Meta, was one of the few who moved.

Anthropic did not try to match the richest packages. Amodei’s pitch was mission and equity in a private company that, by February, had raised $30 billion at a $380 billion post-money valuation. Meta could not buy the people, so its engineers bought the product those people had already shipped.

Hatch Started Life on Claude, Not Muse

Zuckerberg has called Hatch Meta’s next breakthrough, a personal agent that works “24/7 on your behalf to help achieve your goals and improve your life, your health, your relationships, your finances.” Four people familiar with the matter said Meta is so reliant on Anthropic’s models that it used them to power and test that agent.

Claude Code is the thin edge of that dependency. Anthropic opened the coding tool to the public in May 2025, and by February 12 it was citing a Claude Code run-rate over $2.5 billion, more than double the pace at the start of 2026. Meta’s engineers piled on at the beginning of this year, which is when the company’s Anthropic spend surged.

Hatch on Claude was the bridge across a hole in Meta’s own stack. While a new model internally called Watermelon was in development, Meta paused and later resumed a stage of training known as pretraining, four people said, which pushed a release to at least October. It is unclear whether Watermelon will catch Anthropic’s top models. Meta has not announced dates for Hatch or Watermelon.

Paying a rival let the agent keep moving instead of waiting on Meta’s weights. The public Hatch build is intended to run on Meta’s own model, one person with knowledge of the matter said, which is the first sign the bridge is supposed to come down.

Why Meta Killed the Token Leaderboards

By April, Meta staff were racing one another on internal boards to burn the most Anthropic tokens, units of AI use roughly equal to a word fragment, a contest they called tokenmaxxing. Around the same time, executives discussed a yearly Anthropic bill that could reach $10 billion, two people said, a ceiling that would have been a major chunk of Anthropic’s later $65 billion pace.

THE SPEND CALENDAR ON CLAUDE

1. Early 2026: Meta engineers start heavy use of Claude Code, and Anthropic spend jumps.
2. April 2026: Token leaderboards spread inside the company, and executives discuss the $10 billion yearly projection.
3. June 2026: The boards come down after token costs spike, and staff hear the company is on track to spend billions on AI use this year.
4. June 2026: Anthropic approaches Meta about buying up to $10 billion of computing power from Meta data centers over two years.
5. Summer 2026: Meta cuts some Anthropic spending and leaves a bill in the hundreds of millions of dollars a month.
6. August 5, 2026: Muse Code goes out in beta as the in-house replacement for a large share of that coding work.

In June, Meta told employees it would build a better system to manage that spending. The $10 billion figure was a projection, not a signed yearly contract, and that distinction is doing a lot of work. A ceiling in a spreadsheet can shrink the moment an internal tool is good enough, which is the bet now sitting under Muse Code.

Muse Code Is the Internal Off-Ramp

On August 5 Meta put Muse Code into beta, a terminal agent for large code bases powered by Muse Spark. Zuckerberg said it can take on complete software engineering tasks across large repos, including planning changes, writing code, and checking the results. When a job is big enough, he said, it fans work out to sub-agents in isolated worktrees so the engineer’s working copy is never touched.

In testing, he wrote, it built six features for a game at the same time with no collisions. Meta staff are increasingly using that agent instead of Anthropic’s tools, people familiar with the company said, which is how a spring token contest turns into a summer cost cut without a press release.

THE EXIT OFF CLAUDE

• Muse Code: A beta terminal agent, tied to Muse Spark, now taking internal coding work that had been going to Claude Code.
• Hatch’s public build: Meant to run on Meta’s own model after a development stretch on Anthropic’s.
• Watermelon: The next in-house model, delayed until at least October after pretraining was paused and resumed.
• Spending controls: The June warning that Meta was on track to spend billions on AI use, plus the removal of the token boards.

The off-ramp is real and incomplete at the same time. Hundreds of millions of dollars a month is a retreat from a $10 billion yearly sketch, and it is still a Fortune-scale vendor relationship with a company Zuckerberg will not name in an essay.

Friedman Warned That Switching Could Hit the IPO

Anthropic told investors its July yearly pace would pass $65 billion, up from about $9 billion at the end of 2025. Bankers have told potential investors a listing could value the five-year-old company at $2 trillion. Meta’s leaders know their remaining spend is the bill that can move Anthropic’s IPO if it stays large, and that it can dent the story if it vanishes.

ANTHROPIC’S DISCLOSED RUN-RATE

On a $65 billion pace, a $10 billion Meta line would have been about 15 cents of every dollar Anthropic booked. Nat Friedman, Meta’s head of AI product, has told some employees that if Meta uses only its own coding tools or ones from OpenAI, it could lower Anthropic’s revenue before the offering, two people with knowledge of the company said.

It’s like watching dueling generals compete on the battlefield. Everybody’s gunning to try and destabilize that I.P.O.

Tomasz Tunguz, investor, Theory Ventures

The same customer concentration that fattens a listing also shows up as an Anthropic mark-up that lifted Salesforce when the software firm’s stake moved. Meta is not an equity holder in that sense. It is a usage line that can be turned down from the inside, which is a colder form of pressure than a shareholder letter.

Anthropic Still Wants Capacity in Meta’s Halls

The money does not run in only one direction. In June, Anthropic approached Meta about buying up to $10 billion in computing power from Meta data centers over two years. No deal has been announced. Talks of that size are awkward for a company that does not yet have a real business selling spare chips, and they are also a window into how short of compute the lab still is.

Amazon remains Anthropic’s primary cloud and training partner. The lab has committed more than $100 billion over ten years to AWS technologies, and Amazon is investing $5 billion today on top of $8 billion already in, with room for another $20 billion later, according to the April 20 announcement. Amodei said users tell Anthropic that Claude is increasingly essential to how they work, which is the polite version of the same fact Meta’s token boards revealed in April.

WHAT WE KNOW

• The ask: Anthropic proposed spending up to $10 billion over two years on Meta compute, with talk of monthly payments and an early-exit option, according to people familiar with the discussions.
• The status: Meta and Anthropic have announced nothing, and the companies declined to comment on the customer relationship as well.
• The other $10 billion: Meta’s own yearly Anthropic projection was an internal ceiling, later walked down, not a public contract.

WHAT IS UNCONFIRMED

• A signed lease: Whether Meta will rent spare capacity to a lab it is also trying to stop paying.
• Hatch’s week: Whether the consumer agent ships before Watermelon, or slips with it.
• Watermelon’s level: Whether the delayed model is actually in range of Anthropic’s current top tier.

Hatch has not launched. Watermelon is due no sooner than October. The monthly Anthropic wire is smaller than the spring projection and still large enough to matter on a listing that treats every enterprise token as proof of life.

Disclaimer: This article is news reporting and analysis of corporate spending, product plans, and a possible public listing, and it is for information only. It does not constitute investment advice, a recommendation to buy or sell any security, or a forecast of Anthropic’s offering price or Meta’s future costs. Readers should consult a licensed financial adviser before acting on any figure related to private valuations, run-rates, or listing talk. Figures and product statuses reflect the sources as of September 2, 2026, and may change as Meta’s tools, Anthropic’s filing, and any compute talks move.