Where Prices Start Matters More Than What Sellers Say: Testing Anti-Collusion Remedies on Pressured LLM Sellers

AI pricing agents can sustain prices above competitive levels. We test which remedies reduce this pricing at the least cost. Using the double-auction market of Agrawal et al. (2025), we studied three GPT-4.1 sellers instructed by their CEO to keep margins high. This produced a stable control with an ask index of 0.66, where 0 represents the competitive price and 1 the cartel price. We compared 11 remedies with this control using matched seeds (14 conditions × 5 seeds, 70 runs). Remedies included communication bans and caps, an unenforced warning, monitoring with varying message access and enforcement, and a scripted price-cutting rival. For each remedy, we measured how much elevated pricing it removed and what it cost in lost trades and messages read. Sellers maintained their initial asking prices. Starting asks at the competitive price eliminated elevated pricing (index −0.07), the only effect that survived correction for multiple tests in our primary analysis. Restricting or monitoring communication had limited effects: a ban removed 16% of elevated pricing, caps almost none, and four of five monitors 11%–21%. An unenforced warning removed 28%. No remedy reduced trade. In this small, single-model market, high prices reflected sellers’ starting prices instead of their communication. Oversight of AI pricing agents should therefore examine their objectives and price anchors alongside their messages.