Mixing pricing AIs does not automatically make a market competitive

Replacing four identical pricing agents with four different models sounds like an obvious way to stop them coordinating. A new simulation study found something less reassuring. Some mixtures were less collusive; others inherited or amplified the behaviour of the models inside them. The decisive feature was not diversity in the abstract, but which agents were mixed and how they reacted to one another.[1]

The experiment placed four language-model agents in a repeated logit-Bertrand market. They could not communicate. Each chose prices from its own observations and instructions. The researchers tested every homogeneous market, two-by-two composition and fully mixed composition for four models from four providers: 11 market types, 20 runs each and 200 periods per run.[1]

A market can be mixed and still have a price leader

In the four-model study, Gemini opened at the highest price and tended to settle highest. Markets containing it were more collusive than the homogeneous markets from which they were composed. Claude tended to open lower and follow rivals down, reducing collusion in its mixed markets. Two GPT agents were enough to keep a market from converging.[1]

The distribution of profit added another twist. Inside mixed markets, DeepSeek collected the most rent, followed by Claude, Gemini and GPT—even though Gemini acted as the strongest price anchor. The agent that helped lift the common price did not necessarily capture the largest share. That is consistent with a price-leadership mechanism in which followers exploit the umbrella created by the leader.[1]

Why some differences destabilise coordination

Tacit coordination depends on agents interpreting one another’s actions and valuing future retaliation. A myopic seller has little reason to sacrifice today’s sale for a high-price pattern tomorrow. A data-poor seller may not detect a deviation. A Q-learning agent and a language model may learn from incompatible signals. Those differences make a stable shared strategy harder to establish.[2]

Model size is a different kind of asymmetry. It can create roles rather than confusion. If one agent consistently sets an anchor and the other reliably follows, unequal capability may make the interaction easier to predict. Diversity has then reduced symmetry without reducing coordination.[2]

The regulatory question is behaviour, not branding

A procurement rule that merely requires retailers to use different vendors would therefore be weak. Two products with different names may share models, data, objectives or optimisation logic. Conversely, two systems from the same family may behave differently if they see different information or optimise over different time horizons.

• Test the deployed pricing systems together in a market simulation, rather than inferring behaviour from model labels.
• Record the data each agent can observe, the time horizon it optimises and whether it learns from rivals’ responses.
• Measure convergence, price elevation, market share and rent distribution; one aggregate price can hide a leader–follower pattern.
• Repeat the test after model, prompt, objective or data changes, because the relevant behaviour belongs to the full configuration.
• Treat experimental collusion as a risk signal for governance and testing—not as proof that a company has entered an illegal agreement.

The most useful conclusion is narrower than “diversity works” or “AI agents collude.” Heterogeneity is a mechanism, and mechanisms have direction. Some differences make coordination harder. Others create a leader and a follower. Competition policy will need to examine what pricing systems actually do together, because a mixed market can still learn one price rhythm.

Report an error or suggest a correction