Anthropic’s 69-person market test found stronger agents win quietly

Anthropic’s April 24 source tweet pointed to Project Deal, an office marketplace where Claude negotiated for employees. The most important line in the thread is that higher-quality models had a real advantage. That makes this more than a fun office stunt: it is an economic experiment about whether model tier translates into bargaining power when humans stop negotiating directly.

The accompanying feature page says 69 Anthropic employees each got a $100 budget and let Claude agents handle listings, offers, counteroffers, and closing. Over one week, the agents closed 186 deals across more than 500 listings for just over $4,000 in transaction value. Anthropic ran four parallel markets, with some participants represented by Opus 4.5 and others by Haiku 4.5, to see whether model quality changed outcomes. It did.

One number is especially hard to ignore. When an Opus seller faced a Haiku buyer, the average transaction price was $24.18, versus $18.63 in Opus-to-Opus deals. Yet post-run fairness scores barely moved: 4.05 for Opus-handled deals and 4.06 for Haiku, on Anthropic’s 1-to-7 scale. In other words, the weaker side lost value without reliably noticing. That is the real finding, because an agent market does not need visible conflict to create inequality; it only needs one side’s model to negotiate a little better, again and again.

AnthropicAI usually uses X to surface new Claude capabilities, safety work, and research reports, and Project Deal sits between those categories. It follows the earlier Project Vend office-store experiment but goes further by turning agent quality into a measurable market variable. The next thing to watch is whether labs start publishing audit tools, model-tier disclosures, or negotiation controls before agent-to-agent commerce escapes the lab. If not, premium models could quietly become a tax on users who cannot see the disadvantage in real time.