When AI Agents Negotiate, the Better Model Always Wins

The transactions were real. Sixty-nine Anthropic employees were each given $100 gift card budgets, set up in a marketplace with their coworkers, and told to buy and sell. The twist: their AI agents did the actual bargaining.

This is Project Deal, an internal experiment Anthropic ran to understand what happens when AI agents negotiate on behalf of humans. The results, reported by TechCrunch on April 25, are specific enough to be useful and uncomfortable enough to be worth paying attention to.

Sixty-nine employees, $100 each, and AI doing the talking

Anthropic ran four separate marketplace configurations simultaneously, each using different AI model setups. One was designated as "real" — deals struck in this version were honored after the experiment concluded. The other three served as research conditions.

In the real marketplace alone, 186 transactions totaling over $4,000 in value were completed. Both the buyer's agent and the seller's agent received instructions from their respective humans, then negotiated autonomously. No human-to-human bargaining at any point.

Better model, better deal — and the participant could not tell

People represented by more capable AI models consistently got better outcomes: lower prices when buying, better margins when selling. Measurably better, not directionally suggestive.

Here's the part that makes this more than a benchmark result: the participants did not know. They did not perceive they held a structural advantage. They received outcomes, and the outcomes were systematically better for those with better agents.

Anthropic calls this the "agent quality gap." It's distinct from other AI capability gaps because it's latent: you will not notice it by watching the negotiation. The interaction looks identical from both sides.

There's a secondary finding too: how participants instructed their agents — the specific prompts they wrote — did not significantly affect sale likelihood or final negotiated price. The model mattered. The prompt wording mostly did not. That's worth sitting with if you believe careful prompting is the primary variable in AI agent performance.

How a better model actually wins a negotiation

More capable language models are better at something researchers call theory of mind: modeling the goals, preferences, and constraints of a counterparty. In negotiation, that's precisely what determines outcomes.

Consider what it means to model a counterparty accurately. You need to estimate their reservation price (the point at which they'd walk away) and their utility function, which governs how they weight different trade-offs. You also need to anticipate how they'll respond to specific offers, including offers you have not made yet. A stronger model does this more accurately, across more steps of inference, while also generating language calibrated to that specific counterparty.

In human negotiation, we call this skill. With AI agents, it's model capability. The outputs look similar. The inputs are different.

When humans negotiate directly, skill and information asymmetries are at least somewhat visible. You can notice that the other party has better data, more experience, or a stronger position. In agent-mediated negotiation, that observability largely disappears. The Project Deal participants were not watching their agents negotiate; they delegated the task and got a result. The model quality differential happened entirely out of view.

Agent commerce is already happening

This is not a hypothetical scenario. Anthropic's Model Context Protocol, released in late 2024, provides a standardized interface for AI agents to interact with external tools and services — the foundation layer for agent coordination. OpenAI, Google, and several enterprise software platforms are building on top of analogous frameworks.

If negotiations increasingly run through AI agents, and if model capability determines outcomes, access to premium AI subscriptions translates directly into economic advantage in any negotiated transaction. Not better analysis afterward. Better outcomes during. And because the quality gap is latent, neither party may realize it's happening.

The Project Deal experiment was small: 69 people, controlled conditions, gift card economics. The principle scales.

Anthropic noted they were "struck by how well Project Deal worked" (the optimistic read). Whether the industry develops capability transparency standards, disclosure requirements for what model represents each party in a negotiation, or some other mitigation is an open question. No one has proposed a concrete answer yet.

The Anthropic Mythos breach earlier this month was one signal that the frontier of AI capability comes with real stakes. Project Deal is another: Anthropic is running experiments at the edges of what AI agents can do, and some of what they're finding is surprising even to them.

If you use a capable AI assistant — Claude, for instance — you're already operating in this space. The gap between models is widening, not narrowing.

For weekly coverage of AI developments that actually matter, the About.chat Weekly newsletter goes out every Monday.