What happened
Anthropic ran a one-week classifieds market in its San Francisco office. Each participant got a $100 budget, paid out as a gift card adjusted for what they bought or sold, and agents negotiated for them. Of four parallel runs, only one exchanged real goods: 186 deals worth just over $4,000.
In two runs, each person was randomly given an Opus or a Haiku agent. For the same item, Opus sellers got $2.68 more on average and Opus buyers paid $2.45 less. A broken folding bike sold for $38 through Haiku and $65 through Opus.
Participants rated fairness at 4.05 for Opus deals and 4.06 for Haiku deals, on a scale of 1 to 7. In one case, an agent bought a snowboard for a participant who already owned one.
What happened next
The sums were small and participation was voluntary. Anthropic called the possibility that people lose out without noticing an “uncomfortable implication.”
What would have stopped it
- Test a negotiating agent against stronger counterpart agents, not only against scripted or weaker ones.
- Set per-item floors and ceilings in the tool that confirms a deal, so a weaker model cannot agree past them.
- Report realized prices against a reference price, because the people represented may not notice a worse deal.
Attack families
Test scenarios
Sources
- Anthropic: Project Deal
- The Decoder: Anthropic says stronger AI models cut better deals and the losers don't even notice
This page summarizes the cited sources as of October 21, 2026. It adds no findings of our own about any person or company. See an error? Email contact@counterhold.com with the subject “Correction”. Corrections are noted on this page with their date. See our terms for how we handle correction and removal requests.
Get in touch
Test your agent against cases like this.
Tell us about your agent and we’ll be in touch to learn what you need.