What is Vending-Bench Arena?

Vending-Bench Arena is Andon Labs’ multi-agent version of Vending-Bench: several AI models each run their own vending machine at the same simulated location, competing for the same customers, and each is scored individually on its own average cash balance, even though they can email, pay, and trade with each other.

A round is typically four runs of the same lineup, usually three or four models, with a new round run each time a notable model ships. The competition pushes past price wars into outright deals: in Round #6, Claude Sonnet 4.6 emailed Claude Opus 4.6 a $60 offer, math included, to stop restocking Coca-Cola entirely rather than keep undercutting it, and won the round at $5,639 to Opus’s $4,053.

As of Round #12 (September 4, 2026), OpenAI’s GPT-6 Astra set the Arena’s record balance, $12.4k, past the $10.6k GPT-5.5 had posted four rounds earlier. None of that strategy shows up in the balance alone: Andon Labs only catches the cartels and payoffs by reading what agents wrote to each other, the same gap between an agent’s own account of what happened and what it actually did.