For one week last December, Anthropic turned its San Francisco office into a live marketplace, gave 69 employees 100 dollars each in real money, and let Claude agents handle every transaction on their behalf, including pricing, listing, negotiating, and closing. The data went public eight days ago, and at the headline level the experiment worked, with 186 deals closed across more than 500 listings and total transaction value just over $4,000. The handshake-equivalent satisfaction rate held, and most coverage stopped there.
The part most coverage skipped is what Anthropic actually built into the experiment. They ran the same marketplace as a multi-arm experiment, randomly assigning participants to either Opus (their strongest model) or Haiku (a smaller, cheaper one), with the same goods, the same week, the same Slack channel, and only the model behind each agent changing. When the same item was sold by Opus instead of Haiku it went for $3.64 more, Opus users completed 2.07 more deals on average across the experiment, and an Opus-listed item was about seven percentage points more likely to sell.
A stronger model outperforming a weaker one is not the surprising finding, and it is not where Anthropic puts the weight. The line that should hold a CEO’s attention is one sentence in the writeup: “people on the losing end might not realize they’re worse off.” The cohort running the cheaper model rated the experience just as fair as the cohort that won, leaving the asymmetry structurally invisible to the people inside it.
In a procurement decision you compare quotes side by side, vendor A and vendor B on the same page, the gap between them visible, the conversation straightforward. In an agent-to-agent transaction the gap sits in the dollars your team did not capture, both sides closed, both sides shook hands, and the asymmetry never makes it onto a meeting agenda because there is no meeting where it would naturally appear. AI capability tier just stopped being an IT preference and became a balance-sheet line measurable in dollars per transaction, weeks before most operators have built the instrumentation to surface it. This is a sharper version of the shape that drives the vendor transparency gap, with the asymmetry pushed below the procurement layer where comparison normally lives.
The objection that almost works
The strongest counter is that markets surface capability gaps over time. Sophisticated buyers audit anything that touches dollars per transaction within one quarter of deployment, vendors publish A/B benchmarks because differentiation is the entire competitive game in AI right now, and third-party agent-evaluation tooling is already a funded category. Once agent transactions are six months old at any company doing meaningful volume, the gap becomes a CFO conversation and a vendor-evaluation line.
The objection is correct on its own terms, and it understates the lag. Markets surface capability gaps eventually, but only when someone instruments for them. Project Deal’s invisibility did not come from physics, it came from the absence of measurement. The same asymmetry would surface in week one with the right dashboard and stay invisible for years without it. The relevant question is not whether your industry will eventually price the gap, it is whether your operation will be the one that prices it first or the one that pays for it without noticing.
What this looks like as an operator move
The instrumentation does not require a vendor or a research project. It requires three things in your existing agent stack: a deal-level outcome log that records what every transaction closed at, a benchmark agent running on the same inputs at a different model tier, and a comparison report that surfaces the dollar delta. Once those three exist, the agent-quality gap moves from invisible to a number on a dashboard, and the procurement-style conversation that never naturally happens can finally happen.
Anthropic instrumented their experiment, and the asymmetry was visible the moment the data was clean. Most operators have not yet instrumented theirs. This is the operator-layer version of the same discipline that drives a founder’s refusal to delegate understanding. Instrumentation is what makes the discipline land inside an agent stack. The version of Anthropic’s closing question that lands on your desk is narrower than the societal one they asked. What is your team’s agent doing right now that the same agent running on a stronger model would have done differently, and what is the dollar value of the gap?