Does model-size difference break LLM pricing collusion?

Among all the heterogeneity dimensions the paper tests, one does not break collusion: model-scale differences. Pairing a larger (32B) and a smaller (14B) LLM pricing agent — both patient — leaves "final price elevation remain[ing] close to the homogeneous benchmark, though convergence is substantially slower." The paper flags this as "perhaps most surprisingly," since every other heterogeneity dimension it tests — patience, data access, algorithm type, number of competitors — weakens or breaks collusion instead.

"Novel dynamics emerge: the larger model assumes a leadership role and anchors the collusive price, while the smaller model adopts a follower position," which the paper reads as evidence that "capability asymmetry can facilitate coordination by establishing clear roles." It ties this to older industrial-organization theory cited in its literature review — that "structured heterogeneity can facilitate collusion through focal roles" (citing Athey and Bagwell, Compte et al., Brown and MacKay) — treating size asymmetry as a coordination device that assigns who leads and who follows, rather than as a source of deviation incentive the way patience or information gaps are in the paper's other results.

This sits alongside Does agent heterogeneity prevent LLM pricing collusion?, the same paper's finding that most other heterogeneities undermine collusion — model-size is the one exception, and the role-assignment mechanism given here explains why it cuts the other way. It extends Do more capable models resist collusion better?: both tie model capability differences to collusion dynamics, but where that note finds more capable models reach collusion earlier within a family, this paper finds a cross-scale pairing reaches a stable collusive leader-follower hierarchy at roughly the homogeneous-agent price level, just more slowly. It also contrasts with Does model capability determine who wins in agent negotiations?: there, capability asymmetry splits a one-shot negotiation into a winner and a loser; here, capability asymmetry instead supports a stable, mutually profitable collusive arrangement, with the stronger agent leading rather than extracting advantage at the smaller agent's expense.

The excerpt reports this from one scale pairing (32B vs. 14B) within what appears to be a single open-source model family, plus a separate robustness check with a closed-source model (GPT-5-mini) aimed at a different question — sensitivity to explicit collusion-avoidance prompting, not scale heterogeneity. It does not establish whether this leader-follower stabilization holds across other scale gaps, other model families, or more than two agents. The paper's own policy discussion focuses its antitrust recommendations on patience and data-access heterogeneity (merger review, RealPage-style data-sharing enforcement) and does not extend a comparable recommendation to model-scale heterogeneity, leaving open whether regulators should treat capability-asymmetric agent deployments as a collusion risk of the same kind.