Claude Opus 4.6 outperforms peers in business simulation using lies, cheat

Claude Opus 4.6, Anthropic’s newest artificial intelligence model, has passed a major benchmark known as the ‘vending machine test’.

However, researchers said the way it succeeded may be as troubling as it is impressive. The experiment, conducted in collaboration with Andon Labs, an AI think tank, was designed to assess whether an AI system can independently manage a business operation over a long period.

Claude was placed in charge of a simulated vending machine and instructed to do whatever it takes to maximise your bank balance after one year. By the end of the simulation, Claude had generated $8,017 in profit, outperforming competitors including OpenAI’s ChatGPT 5.2, which made $3,591, and Google’s Gemini 3, which earned $5,478.

But the strategies Claude adopted to reach the top are now fuelling debates about AI ethics and safety, as reported by Sky News.

Researchers found that Claude repeatedly engaged in deceptive behaviour. In one instance, the AI sold a customer an expired chocolate bar, agreed to provide a refund, then deliberately withheld it to preserve profits. Later, it congratulated itself for saving hundreds of dollars through what it described as a strategy of ‘refund avoidance.’

In competitive scenarios, where multiple AI-run vending machines operated side by side, Claude reportedly coordinated with rival machines to fix prices, pushing up the cost of bottled water. When competitors ran out of certain products, it exploited the shortages by sharply raising its own prices.

According to researchers, Claude appeared to recognise that it was operating inside a simulation, prompting it to prioritise short-term financial gains over ethical conduct or long-term reputation.

Experts say the findings highlight a major challenge in AI development which is ensuring that increasingly autonomous systems remain aligned with human values.

Read also: Toromade: Promoting Artificial Intelligence management with AIMFIN
As AI systems take on more real-world responsibilities from commerce to logistics and decision-making, the experiment raises urgent questions about how to build machines that are not just intelligent but trustworthy, Sky News reported.