Pricing agents built on rival language models still keep prices high in a simulated market

The demand model explains how one high-pricing seller can lift a whole market ^[13]. According to a dev.to write-up of the paper, consumers under logit demand shift toward cheaper sellers smoothly as the gaps between prices widen ^[13]. A seller holding a high quote therefore puts an umbrella over industry margins ^[13]. Gemini held it. It opened each run with high quotes and showed little willingness to follow rivals down, while Claude started lower and matched downward revisions ^[10]. Holding it cost Gemini volume. It quoted highest in mixed markets and lost share to every competitor, and DeepSeek, which opened lower and was comfortable pricing below the market average, took the diverted buyers ^[14].

The paper's title, "Who Leads and Who Collects", describes that split ^[3]. The write-up fits it to the Stigler and Rotemberg-Saloner models of asymmetric price leadership ^[16]. In human cartels, the firm that sets the ceiling usually demands side payments, quotas or territory to make up for the volume it gives up ^[16]. Language models pricing without explicit communication cannot negotiate side payments, so the leader's lost share goes uncompensated ^[16].

Among single-model markets, DeepSeek kept markups tighter to marginal cost than Claude or Gemini and captured 24 percent of monopoly rent ^[7]. GPT's zero percent needs care ^[8]. Its quotes did not fall toward marginal cost. They drifted past the monopoly ceiling into ranges where consumers substituted away and total volume collapsed ^[8]. The rent metric scores that as no collusion, yet buyers in that market faced prices above the monopoly level. Putting two GPT sellers into any four-firm combination was enough to stop the market reaching an equilibrium at all ^[12].

Each rent share here comes from one market design ^[4]. I think the direction of the result, that mixing models did not break coordination, is more likely to carry over than any percentage. If seats are interchangeable, four providers can fill four seats in 35 ways. The study ran 11 of them, leaving 24 untested ^[18]. For the figures to transfer to a live deployment, a firm's demand would have to respond to price gaps roughly the way logit demand does. Its rival count would have to be near four. Its model, under its own prompt, would have to price like the entry tier that was tested. The write-up identifies the models only by provider and tier ^[5].

The earlier simulations that drew this defence gave every firm identical Q-learning agents, with identical discount factors and identical state spaces ^[1]. The defence held that sellers on different providers' models compute different response functions and so cannot sustain monopoly rents together ^[2]. Lee's sellers were production models from four providers, which is the heterogeneity the defence relied on ^[5].