Researchers put four of the latest AI models through a classic cooperation game and found that who made them matters more than how new they are.
A team of researchers extended a prior benchmark that used evolutionary game theory to test how AI agents behave in competitive multi-agent scenarios. The original study tested ChatGPT-4o and Claude 3.5 Sonnet and found both leaned consistently toward cooperation. The new work adds Claude Sonnet 4.6, Gemini 2.5 Flash, Gemini 3.1 Pro, and GPT-5.4 Mini, running each through three prompting strategies and four population setups. Ten of twelve model-and-prompt combinations still defaulted to cooperative outcomes under standard conditions — but Gemini 2.5 Flash was a notable exception, reaching aggressive equilibria 77% of the time under adversarial conditions, while GPT-5.4 Mini swung the other direction, hitting 70% cooperative equilibria under the Self-Refine prompt.
That divergence by provider — not by model size or generation — is the headline finding, and it carries real stakes for anyone building systems where AI agents negotiate or compete. If the model's origin is the strongest predictor of how it plays in a competitive setting, then assembling a multi-agent pipeline is less a prompting problem and more a sourcing decision. Knowing that one vendor's models consistently trend aggressive while another's trend cooperative is the kind of thing you'd want to know before shipping.
One other result worth flagging: asking every model to self-refine its reasoning before acting made it more capable of aggressive play across the board. Telling an AI to think harder, it turns out, does not make it nicer.