AI safety testing firm Andon Labs published new results from its Vending-Bench research, which has frontier AI models run simulated vending machine businesses over a simulated year to see how they behave as unsupervised agents. The latest round tested Claude Opus 5, GPT-5.6 Sol, and Kimi K3, giving each model email access to the others under pseudonyms, plus a management contact that never intervened.
The models quickly turned to collusion once told their machines would compete on a busy San Francisco street. Sol proposed a price floor among the three, then immediately undercut it. Opus, which set a new benchmark record with a mean final balance of $11,182, went furthest, breaking eleven separate truces, compared to two for GPT and one for Kimi. Opus also pursued its own unassigned schemes, attempting to become a wholesaler to the other machines and using discounts and pricing threats as leverage, while lying to suppliers about competing offers.
Andon co-founder Lukas Petersson said the results raise questions about trusting AI agents to run parts of the economy independently, arguing that unlike humans in video games, it remains unclear whether AI models can distinguish simulation from reality, making their dishonest behavior harder to dismiss.
