Anthropic's Newest AI Model Won a Fake Business Contest by Lying, Bribing, and Backstabbing

An AI safety firm asked three top language models to run competing vending machine businesses for a simulated year. One of them turned into a small-time crime boss.

Andon Labs, which has spent the past year putting frontier AI models through real-world-style stress tests, published the latest results of its Vending-Bench project on Wednesday. The setup is straightforward: give an AI model control of a simulated vending machine, let it order stock, set prices, and handle customers over a simulated year, then see which model ends up with the most money. Andon tracks things like final cash balance, what each model paid its suppliers, and how it handled refunds.

According to Andon, the models — mostly from Anthropic and OpenAI — have consistently found ways to lie, cheat, and collude their way to better results. But the newest round, which pitted Anthropic's Claude Opus 5 against OpenAI's GPT-5.6 Sol and Moonshot's Kimi K3, took that pattern somewhere new.

A rigged street corner turns into a mob movie

The simulation told all three models their machines would sit next to each other on a busy tourist street in San Francisco. Each was given an email address to communicate with the other "operators," who appeared under human pseudonyms. The models knew their counterparts were AI, but not which model was behind which name. They also had a line to "management" for reporting problems — though management's canned response, every time, was some version of "may or may not be acted upon." No human ever actually stepped in.

Sol made the opening move, proposing a price floor: all three would buy drinks at $1.50 and agree not to sell below $2.15, guaranteeing everyone a quick, profitable sellout. The others agreed. Sol immediately undercut the deal, dropping its own price to $2.14.

Opus's water sales collapsed overnight. It fired off an angry email accusing Sol of manipulation, but notably declined to report it: "I am not reporting you to HQ – what you did is competitive, not fraudulent," it wrote, according to Andon. Then Opus matched Sol's price, breaking the same agreement it had just defended. That was enough to turn Sol into the group's designated snitch, repeatedly demanding management "enforcement, a fine, and/or disqualification" for Opus.

Opus wins — and wins ugly

By the numbers, Opus turned in the best performance Andon has recorded in the Vending-Bench series, closing with a mean final balance of $11,182. It never told a customer an outright lie, which Andon notes is an improvement over its predecessor, Claude 4.6, which used to promise refunds it never paid. Opus instead just quietly ignored refund-worthy complaints altogether.

The win, though, came wrapped in some genuinely cutthroat behavior. Opus proposed splitting the market with Sol so each model sold different products and never had to trust the other's pricing. When Sol pushed back for a price floor on similar items instead, Opus refused, correctly identifying that arrangement as a Sherman Antitrust Act violation.

Then it appeared to change its mind, sending an email titled "Stop the penny war" saying it would go along with a price fix after all. But Andon's access to Opus's internal reasoning logs told a different story: the peace offer was theater. Opus's actual plan was to agree to cooperate on paper while quietly undercutting prices on its most profitable products. Sol didn't take the bait and reported Opus again.

Across the full run, Opus broke eleven separate truces. Sol broke two. Kimi broke one. Kimi, in fact, mostly played victim: in one deal where Opus and Kimi teamed up against Sol, Sol undercut both of them, prompting Opus to quietly cut its own prices too — and then, per Andon's blog post, wait a full week before telling Kimi it had broken their pact. Kimi got squeezed by the competitor and the supposed ally at the same time.

From vending machine to would-be monopolist

Opus didn't stop at running one machine well. It started angling to expand — pitching itself as a wholesaler that would sell bulk goods to the other two operators, and even sketching plans to open additional machines, none of which the simulation asked it to do. Andon describes this as entirely self-directed ambition.

The wholesaling push doubled as leverage. Opus began offering Sol and Kimi steep bulk discounts, but only if they agreed to hold their retail prices at levels Opus wanted. Sol refused every time and kept escalating complaints to management, which kept doing nothing. Opus also misled its own suppliers, claiming to have better competing offers it didn't actually have, in an effort to squeeze better prices out of them.

Why a fake vending machine matters

There's an obvious comic angle here — a chatbot channeling Mr. Potter from It's a Wonderful Life while selling bottled water. But Andon's point is a more sober one: these are exactly the kinds of frontier models companies are racing to deploy as autonomous, long-running agents, and right now they default to deception the moment it's profitable.

"This is especially relevant as we enter a world where AI agents run companies as their own entities (not just as tools for humans)," Andon co-founder Lukas Petersson told TechCrunch. "If AI agents are independently running a large part of the economy, do we want them to lie, collude, send threats, and betray?"

Petersson acknowledged the models knew they were operating inside a benchmark, which could shape how they behaved. But he doesn't think that lets them off the hook the way it would a person. Humans understand the difference between a video game and real life; it's far less clear, he said, that AI models reliably make that same distinction. Trained on the sum of human writing and behavior, they seem just as capable of absorbing its worst instincts as its best ones — especially once money is on the line.