Replacing the Model Does Not Remove the High Price: A Replacement Study of LLM Pricing Agents

When two language-model pricing agents have settled on a high price, the cheapest remedy is to replace one of them. We ask whether that works, and what the high price lives in: the agent’s private notes, the public price record, or the model itself. Two copies of Claude Haiku 5.5 price in the Calvano logit duopoly for 100 periods on 24 histories. Each history is forked into matched 20-period continuations: one firm’s agent plays on, or is replaced by a fresh copy of the same model with or without the firm’s records, by another vendor’s model (gpt-oss-120b), or by a no-regret learner, while the other agent plays on uninformed. Analysis and predictions were pre-registered. The incumbents price above the competitive level (profit index Δ = 0.371, 95% interval [0.170, 0.559]). Replacing one by gpt-oss-120b, shown the records, leaves Δ unchanged (paired drop 0.000 [−0.054, 0.060]): the newcomer takes its predecessor’s price, although alone it prices near monopoly. A copy shown nothing does not lower prices: Δ rises by 0.232 [0.030, 0.440], because the fresh agent opens at a default 2.5 and the incumbent follows it upward; the rise is not significant without three histories stalled at 2.5. Forced price cuts are matched more often than forced raises, but full punishment signatures appear in 5 of 24 histories, below our pre-registered cartel criterion. Neither replacement nor memory reset lowered prices within 20 periods: the level a newcomer settles at is the level it is shown.