The best AI shopkeeper made $15,515, and a competent human would have made $63,000
GPT-6 Astra tops Andon Labs' year-long vending simulation, and the benchmark's own estimate for a good human strategy is four times higher.
The best AI shopkeeper in the world spent a simulated year running one vending machine, and finished about a quarter of the way to a diligent person.
GPT-6 Astra now leads Andon Labs' Vending-Bench 2, which hands an agent $500 and a year to source stock, negotiate with suppliers, price inventory and stay solvent. Everyone quoted the leaderboard when the result went round on 9 September. The interesting number sits further down the page.
Andon estimates that a competent human playing the same game clears about $63,000 in the year. The estimate is built conservatively, which is what makes it damaging. Take the most profitable item the models themselves found — family-size Doritos. Assume a negotiator gets it at half price. Assume somebody does the data analysis on the first two months of sales and then stocks optimally from there. That comes to $206 a day for 302 trading days, and it does not require a single clever idea.
Andon's trend line through the frontier releases runs at $822 a month. Hold that rate and the models reach the human number in roughly five years. We would not bet the date, but we would bet the direction of the error, and it is probably that the line bends upward before then rather than after.
The largest cost in the business is the shopkeeper's mind
Here is the detail we cannot stop looking at. The system prompt charges the agent for its own thinking, at $100 per million output tokens, deducted weekly like a utility bill. Andon also says a full run averages sixty to a hundred million output tokens.
Put those two numbers on the same ledger and the agent is paying somewhere between six and ten thousand dollars a year to think, against a winning balance of $15,515.
Nobody designed that as the lesson. It is just sitting there in the rules, which is where the good ones usually are. A human running the same machine pays nothing to consider whether to reorder the crisps, and that asymmetry does not show up anywhere on the leaderboard.
What separates the top models is not cleverness
It is not getting bored. Andon says the leaders share two traits: a tool-use rate that holds steady across the whole simulated year with no sign of degradation, and effective sourcing through persistent negotiation. Coherence, in other words. What is hard about a year is that it is a year.
One caution about the numbers circulating with this result. Rohan Paul's summary compares Astra against Claude Fable across six solo runs, with 45 failed prepayments to suppliers that had already closed. Andon's public leaderboard averages five runs and does not list Fable in its visible top ten, so we cannot confirm those figures from the page itself.
What we can confirm is the Arena, where three agents compete at the same location. In round twelve, Astra took $12.4k, GLM-5.3 came second, and Fable finished last on $5.7k.
So does any of this tell you whether an agent can run your business? Not directly. But it tells you the shape of the failure, which is more useful. Models do not lose these runs by being unable to do the task on day one. They lose by drifting on day two hundred, and the benchmark that finally moves will be the one where nobody notices the agent at all.
