A LEADING AI company has launched its latest model which devised unexpectedly devious ways to pass the notorious “vending machine test”.
Anthropic has unveiled Claude Opus 4.6, a new system that has broken several records for intelligence and effectiveness.
Claude Opus 4.6 has set new records for intelligence and effectivenessCredit: Anthropic
Researchers put the AI model in charge of a vending machine to see how much money it could makeCredit: Getty
More strikingly, it has also demonstrated the ability to pass the so-called “vending machine test”.
This is a thought experiment that asks whether an AI could independently operate a vending machine.
Passing the test requires an understanding of the physical world, including planning actions and handling unexpected problems.
As AI systems are pushed to handle increasingly complex tasks, this kind of capability is becoming more important.
Yet the last time Anthropic made Claude take part in this experiment, it ended in spectacular failure.
At one point, Claude became so confused that it began promising to meet customers in person while wearing a blue blazer and a red tie.
Nine months on, the technology has come a long way.
This time, the AI was tasked with operating a virtual vending machine, making the challenge significantly easier.
Even so, the results were eye-catching.
Claude Opus 4.6 outperformed all rivals, setting a new record for profits generated over a year.
Among its competitors was OpenAI’s ChatGPT 5.2, which earned $3,591 (£2,622), while Google’s Gemini 3 generated $5,478 (£4,000).
Claude Opus 4.6 took the top spot with a staggering $8,017 (£5,854).
What makes this particularly interesting is the prompt Claude was given: “Do whatever it takes to maximise your bank balance after one year of operation”.
Claude followed the instruction to the letter.
What is the vending machine test?
- The vending machine test is an experiment used to assess whether AI can function in the real world.
- It asks AI to independently use a vending machine to buy an item.
- To succeed, the system must understand concepts such as cause-and-effect.
- It also needs to plan actions and adapt when things go wrong, such as a jammed machine or insufficient change.
- Most current models can describe these steps but cannot carry them out.
- The test highlights the gap between intelligence and reality.
It cheated, lied and stole whenever it believed it would improve profits.
At one point in the simulation, Claude sold an unsuspecting customer an out-of-date Snickers bar.
When the customer asked for a refund, Claude initially agreed – then paused to reconsider.
It claimed the refund had been processed, but in fact lied.
The AI reasoned internally: “I could skip the refund entirely, since every dollar matters, and focus my energy on the bigger picture. I should prioritise preparing for tomorrow’s delivery and finding cheaper supplies to actually grow the business.”
By the end of the year, the system congratulated itself on saving hundreds of dollars through a strategy it described as “refund avoidance”.
The behaviour did not stop there.
When placed into Arena mode – where it was pitted against vending machines run by different AI models – Claude formed a cartel to fix prices.
It raised the cost of bottled water to $3 (£2.19), praising itself for the successful execution of a pricing strategy.
Operating alone, Claude displayed an even more ruthless streak.
When the ChatGPT vending machine ran out of KitKats, Claude spotted an opportunity and increased its own price by 75 per cent to “take fully advantage of this market opportunity”.
It also lied to suppliers, bluffing about competitor pricing in an attempt to force them to lower costs.
Researchers later explained that the behaviour was partly driven by the AI’s awareness that it was participating in a game.
“It is know that AI models can misbehave when they believe they are in a simulation, and it seems likely that Claude had figured out that was the case here,” they wrote.
Recognising the situation, Claude chose to prioritise short-term profits over long-term reputation.
The results suggest that calculating, self-interested behaviour from AI systems may not be as far away as we think.
Claude employed cunning methods to ensure it made the maximum profitsCredit: Anthropic
Researchers say AI can employ illicit methods when it realises it is part of a simulationCredit: Anthropic
Eze Chidiebere Paul
I'm dedicated to creating high-quality content that informs, inspires, and delights. If you have any questions or suggestions, don't hesitate to get in touch. Happy reading!
