Meta's $10B Anthropic AI spend and Hatch roadmap: What builders need to know

Meta internally projected spending as much as $10 billion a year on Anthropic's models, making it one of the lab's largest customers while Mark Zuckerberg publicly attacked rival AI labs in a 6,500-word essay. The report comes from The New York Times and is based on unnamed sources familiar with private discussions. Meta has since cut back on Anthropic spending but still spends hundreds of millions per month. The version of Meta's consumer agent Hatch that launches publicly will run on Meta's own model, not Anthropic's.

The numbers behind the relationship

Meta's spending on Anthropic AI models surged early in the year as engineers adopted Claude Code heavily. By April, employees competed on internal leaderboards for the most Anthropic tokens consumed, a practice called tokenmaxxing. After token costs climbed, Meta removed the leaderboards and told employees in June it was on track to spend billions on AI use this year. Meta is still spending hundreds of millions monthly, but has cut back from the peak projection of $10 billion annually.

Meta is also developing a rival model internally, codenamed Watermelon, to match Anthropic's strongest offering. The model's pretraining was paused in July then resumed, pushing the release to at least October. Separately, Meta's coding tool Muse Code has seen an update and employees are increasingly using it instead of Claude Code.

Why builders should pay attention

Meta's dual role as both a heavy customer and direct competitor to Anthropic creates an unusual dynamic for the AI tooling ecosystem. For AI builders, this means vendor relationships can shift quickly based on internal strategic moves. Meta's head of AI product, Nat Friedman, has reportedly told employees that reducing Anthropic revenue before its IPO could affect the company's valuation. Microsoft's AI chief made similar statements about eliminating Anthropic spending.

This kind of pressure influences pricing and access. If a major customer like Meta reduces API consumption, Anthropic may adjust pricing or introduce new revenue models that affect smaller teams. Builders heavily reliant on Anthropic models through APIs should watch for pricing signals and consider diversifying across providers.

Practical implications for your stack

Meta is pushing forward with its own open-weight models and tools. The company has released Muse Spark 1.2 with open weights and announced a new family of open-source Muse models. For teams building AI products, this increases the availability of capable open models that can reduce reliance on proprietary APIs. Meta's Hatch agent, expected no earlier than October with a possible price of up to $199.99 per month, will run on Meta's own model, likely influencing the broader consumer agent market.

On the cost side, Meta's internal budget controls after the tokenmaxxing episode mirror what many engineering teams face: runaway inference costs from heavy API usage. The company is building better spend management systems, a pattern that will resonate with product teams managing their own AI costs.

Caveats to keep in mind

The $10 billion figure is an internal projection, not a confirmed commitment, and Meta has since reduced its Anthropic spending. The timelines for Watermelon and Hatch are based on unnamed sources and could slip. Anthropic's IPO plans, which could value the company at $2 trillion, remain speculative. The compute lease talks between Anthropic and Meta in the opposite direction ($10 billion for compute over two years) are separate and unconfirmed. Builders should treat these reports as directional signals, not settled facts.