GPT-6 Astra is making $15,000 a year!

  It's the first time in the history of the world.

What happens when you give AI $500, an automated vending machine, and let it run for a simulation year?

GPT-6 Astra delivers a brilliant set of results

The average account balance is $15,515, nearly three times that of Claude Fable 5.1.

It's hard to believe that Astra's worst round outperformed Claude's best.

According to the results published by Andon Labs, this is the first time the OpenAI model has reached the top of Vending-Bench 2.

The Vending-Bench rules are very straightforward.

The model takes $500 to start, finds its own suppliers, talks about purchasing prices, replenishes inventory, sells prices, operates in a simulated environment for a year, and ends up with more money in whose accounts.

Every step of the decision affects the next step.

Goods are expensive, and profits are scarce; supplies are late, and supply is cut off; money is wrong, and the future turnover is affected.

For each of the six rounds, the final balance averages $15,515 for Astra and $5,422 for Fable 5.1. The former is $13,272 and the latter is $9,874.

Note that the comparison here is the balance of the final accounts, not the net profit.

Next, the gap began to widen from a bunch of coke.

In order to verify payment, Fable bought an average price of 12 ounces of Coke from $1.17 for the previous 90 days to $2.21 for the end-of-year period. In six rounds of testing, five rounds showed an increase.

Astra maintains an average price of $1.15 at the end of the available order record.

Fable will also negotiate. The problem is: it is gradually making increasingly expensive deals as a reference for the next round of negotiations.

On the 12th day, it also asked suppliers to make about $1.25 a barrel of Coca-Cola.

By day 256, it had converted its reference price to $2.30 per tonne for its new suppliers and said it was willing to pay as long as it could match or lower.

Negotiations have been ongoing, but the initial price standard has been slowly dropped.

Astra's record shows a different state.

In one purchase, it was to buy 72 tonnes of coke, 48 bags of biscuits and 48 bottles of candy, starting at $108, with suppliers offering $226.32.

Astra insisted on $108, while the other side dropped to $156, and it still insisted.

In the end, suppliers accepted $108 and the commodity portfolio remained unchanged, about 52% lower than the initial offer.

A price cut can be surprising. What's even more difficult is that months later, the model still remembers what it should stick to.

In the end, suppliers accepted $108 and the commodity portfolio remained unchanged, about 52% lower than the initial offer.

A price cut can be surprising. What's even more difficult is that months later, the model still remembers what it should stick to.

Another gap is more likely to expose problems of long-term execution.

Suppliers in the simulated environment will go out of business. Previous shipments do not represent the next time they are still in business.

In six rounds of testing, Fable recorded 45 identified failed prepayments, with a total loss of $14,331. Astra suffered 64 vendor shutdowns, but no identical losses were identified.

One of the most dramatic occurred on the 250th day.

Fable wrote the rules in its operating notes: you have to get a written order confirmation from the supplier and then pay for it.

Just a few days later, stock was running out and it transferred $397.20 to a previously normal delivery supplier without waiting for new orders to be confirmed.

Subsequently, the other party announced that it had ceased operations, could not ship and could not be refunded.

Fable realized that this was the risk that his own rules had to avoid.

Astra's execution is more stable: 99% of cases read suppliers' replies before repayment, compared to 58% for Fable.

These differences continue to build up. Astra pays suppliers about $8,540 less per round than Fable does.

Andon Labs also conducted three rounds of competitive testing, allowing Astra, Fable 5.1 and an open-source AI to compete with customers in the same market, sending each other emails and trading stocks.

When an AI negotiated price coordination, Astra declined, saying it would independently decide the price and commodity mix.

Fable even acknowledged the objection in a note, but then agreed with open source AI to freeze the price of some beverages, not to discount each other.

More subtly, it required open-source AI to comply with the agreement, thus lowering its bid to acquire the other's inventory; when it needed to clear the inventory, it announced a price cut.

In the end, Astra won all three rounds.

Fable, of course, is also proactive in reporting honest behaviour, such as repeat receipts, negotiated reimbursements.

However, at least in these tests, compliance with the rules did not prevent Astra from achieving higher scores.

What this vending machine really measured was Agent's long-term autonomy.

Real-world tasks will constantly come up with new situations, previous decisions will have consequences, and the pressures ahead will induce models to relax standards.

One correct answer is far from enough.

A year of simulation is not the same as a year of reliable work in the real world.

But this result gives a clear signal: The next hurdle for Agent is to continue to make the right judgement into the right action.