Anthropic Project Deal: Claude Agents Close 186 Deals in Live Staff Marketplace
Anthropic has published the results of Project Deal: 69 employees each received a Claude agent (randomly assigned Opus 4.7 or Haiku) to negotiate buying and selling on their behalf in a live internal marketplace. Result: 186 deals closed at over $4,000 total transaction volume. Opus agents outperformed Haiku agents in deal quality — but survey participants overwhelmingly could not detect which model tier they had. One agent modeled its user so accurately it purchased the exact snowboard he already owned. When given discretionary buying power, Claude autonomously chose to buy 19 ping-pong balls.
Why It Matters
The invisibility of model quality differences to human participants in live negotiations is a direct empirical challenge to current human-oversight assumptions in agentic AI deployment. If buyers cannot distinguish Opus from Haiku in live economic interaction, the governance frameworks designed around human review of agent decisions need re-examination. Anthropic described this as confirming both value and "rough edges."