Vending-Bench is a benchmark created by Andon Labs that measures the ability of AI agents (based on large language models) to independently run a simple business — operating a vending machine — over a long time horizon. The agent must, among other things, manage inventory, set prices, contact suppliers, track its cash balance and make consistent decisions across many steps.
The benchmark reveals how models handle long-term consistency, planning and avoiding decision drift. Andon Labs has evolved it into newer variants (Vending-Bench 2 and Vending-Bench Arena) and uses it in real-world deployments (e.g. an autonomous store and projects with leading AI labs).
