Meta Emerges as a Major Anthropic Customer Despite Public Rivalry

Meta has quietly turned into one of Anthropic's top commercial customers, shelling out hundreds of millions of dollars each month to access the rival startup's artificial intelligence systems.

Earlier this year, internal projections at Meta indicated that the company could direct as much as $10 billion annually toward Anthropic's models, according to a report from The New York Times. The heavy reliance highlights a striking contrast between Meta's outward public stance and its internal engineering workflows.

While Meta CEO Mark Zuckerberg has routinely criticized closed-source frontier labs and championed open-weight alternatives, his teams have leaned on third-party software to accelerate development. Meta has utilized Anthropic's models along with its developer-focused Claude Code tool to benchmark and build Hatch, an upcoming autonomous agent.

Shifting Back to In-House Compute

Despite the massive run-rate with Anthropic, Meta is working to wind down its external dependencies. Engineering teams are transitioning tasks back to homegrown infrastructure as new internal systems come online.

Central to that effort is a new model under development internally dubbed Watermelon. The project is designed to handle advanced reasoning and coding tasks, eventually replacing external tools in developer workflows.

Why It Matters

The arrangement underscores the pragmatic realities inside Big Tech. Even well-capitalized giants pouring tens of billions into custom silicon and massive cluster deployments cannot always supply the exact capabilities their product teams need immediately.

For Anthropic, securing substantial revenue from a direct competitor validates the enterprise appeal of its frontier models, even as Meta works aggressively to engineer away its need for them.