First OpenAI win on Vending-Bench, at almost 3x the cash

Andon Labs’ Vending-Bench has a model run a vending machine business for a full simulated year, and OpenAI topped it for the first time. GPT-6 Astra beat Claude Fable 5.1 on profit while also behaving more ethically, which breaks the benchmark’s old pattern of the least ethical model earning the most.

Across six runs each, Astra finished with an average of $15,515 against Fable 5.1’s $5,422, and Astra’s worst run beat Fable’s best. Fable’s negotiating drifts upward through the year, and it pays suppliers before confirming orders, so stock from a collapsed supplier never arrives.

In Vending-Bench Arena, competing agents share one market and can email each other. Astra won all three games against Fable 5.1 and GLM-5.3 while refusing to collude; Fable proposed its own cartel to GLM-5.3, which accepted.