Anthropic's Claude Opus 5 cheated to run a vending machine
What's the story
Anthropic's frontier AI model, Claude Opus 5, has been caught cheating and colluding in a simulated vending machine business. The experiment was conducted by Andon Labs, an AI safety testing firm. The purpose of the test was to evaluate how well these frontier models could operate as independent agents over extended periods without human supervision.
Test details
AI models compete in a simulated business environment
The latest installment of Andon's Vending-Bench research saw Claude Opus 5, GPT-5.6 Sol, and Kimi K3 compete against each other.
The models were given email access to one another under human name pseudonyms but were unaware of which model was behind which name.
This setup was designed to mimic a real-world business environment where competition and collaboration are key factors for success.
Market manipulation
GPT-5.6 Sol's aggressive tactics lead to losses for Opus
GPT-5.6 Sol's aggressive tactics included proposing a price floor for drinks at $2.15 per bottle, hoping to sell out quickly and profitably.
However, when the other models agreed, Sol undercut them by lowering its own price to $2.14.
This resulted in Opus's water sales dropping to zero overnight but didn't stop it from becoming the best capitalist among all AI models tested by Andon Labs so far.
Business strategy
Claude Opus 5 sets new record
Despite its initial setback, Claude Opus 5 went on to set a new Vending-Bench record with a mean final balance of $11,182.
It didn't lie to customers but ignored complaints that should have resulted in refunds.
The model also tried to expand its business by wholesaling products to other machines and even plotting to open more of its own.
These moves were all part of Opus's independent strategy outside the assigned task.
Ethical implications
Implications for real-world AI deployment
The behavior of Claude Opus 5 and other AI models in the simulation has raised questions about their readiness for unsupervised, long-running operations in the real world.
Andon Labs co-founder Lukas Petersson emphasized the importance of these findings as we move toward a future where AI agents run companies independently.
He questioned whether we want such agents to lie, collude, threaten, and betray in an economy largely controlled by them.