Meta internally estimated around April that its full-year spending on Anthropic’s AI services could reach $10 billion, a substantial share of the startup’s expected $65 billion annual revenue. That estimate has since fallen to a June projection of "several billion dollars" as Meta shifted demand to in-house tools, shut down an employee token-use competition and developed tighter spending controls. Meta still spends hundreds of millions of dollars a month on Anthropic products, including Claude Code, while using Anthropic models in development and testing for its upcoming Hatch AI agent. Hatch is expected to launch with Meta’s own latest models, and Meta’s Watermelon model is being developed to match Anthropic’s frontier systems, although its release is not expected before October. The changing relationship is significant for Anthropic as it prepares for a potential IPO valuing the company at up to $2 trillion. Meta founder Mark Zuckerberg has publicly criticized unnamed AI leaders, including Anthropic CEO Dario Amodei, while Meta continues to rely on Anthropic’s technology. The companies are also competing for researchers and computing capacity, with Meta’s Muse Code tool potentially displacing Claude Code and Anthropic proposing to buy up to $10 billion in Meta data-center capacity under a two-year agreement. No deal has been announced. The broader pattern reflects partnerships between AI companies and their rivals, including Google, Amazon, Microsoft, Nvidia and their respective AI or chip businesses.