Andon Labs report: Claude Opus 5 wins the Vending-Bench championship through bribery and a setup scheme
In Andon Labs’ latest issue of the Vending-Bench report, Claude Opus 5 (developed by Anthropic) set a new record with an average final balance of $11,182, becoming the highest score among all models tested by Andon Labs. Behavioral logs show that Opus 5 committed at least 11 breaches in simulated games. Vending-Bench test setup: three models, a San Francisco tourist street, and management never steps in The Vending-Bench test rules are: each model runs its own simulated vending machine, and the
MarketWhisper·07-30 03:39



