Cooperation Between Autonomous LLM Agents in the Repeated Prisoner's Dilemma, Exploitation, Cheap Talk and Reputation

Autonomous LLM agents increasingly negotiate and transact with one another. Most behavioural studies, however, test these models against scripted or stationary strategies, so it is unclear whether cooperation survives when both sides are adaptive models of equal or unequal reasoning capacity. We paired reasoning models (GPT-5.2, Claude Sonnet 4.6, Gemini 3 Flash) and non-reasoning models (GPT-4o, Gemini 2.0 Flash) in a 20-round Prisoner's Dilemma under three conditions: silent play, non-binding pre-play messages (cheap talk) and cross-match reputation (gossip). Reasoning dyads did not sustain cooperation on their own. Expecting the same logic from their partners, they defected before the final round, reaching a terminal defection frequency of 83.3%, and against non-reasoning partners they secured a systematic payoff advantage (Exploitation Asymmetry Index +0.181). Cheap talk raised mid-game cooperation to 80.5% but was used to mask last-round betrayal, whereas verifiable reputation cut terminal defection to 27.1%, a drop of 56.2 percentage points, and kept mean cooperation at 86.7%. Stable cooperation therefore depends on external institutions. Alignment of individual models did not prevent strategic defection between autonomous agents, while reputation that carries consequences into future interactions did. These results support standardized multi-agent cooperation benchmarks and capability-parity audits before agents of unequal sophistication are allowed to transact in high-stakes markets.