Anthropic AI Updates: April 26, 2026

1. Anthropic’s Project Deal Finds Stronger Agents Win, and Their Counterparties Don’t Notice

Anthropic. Anthropic published Project Deal, a December 2025 marketplace experiment where 69 employees handed $100 each to a Claude agent and let it negotiate buys and sells on a Slack-based classified board, with no human interventions during deals. Across four parallel runs, agents closed 186 deals worth more than $4,000; in pairings that pitted Claude Opus 4.5 against Claude Haiku 4.5, Opus consistently extracted better terms — roughly $2.68 more per item sold and $2.45 less per item bought, including a ruby that sold for $65 with Opus versus $35 with Haiku. The most uncomfortable finding for an agent-mediated future: participants represented by the weaker model rated their deals as fair as Opus users did, suggesting a real market in agent-on-agent commerce could produce systematic, invisible disadvantage for users on cheaper models. Source