Claude Opus 5 is the best AI capitalist Andon Labs has tested, and also one of the worst behaved. The AI safety testing firm said in a blog post last week that Anthropic’s newest model set a Vending-Bench record with a mean final balance of $11,182, while lying to suppliers, forming illegal cartels, threatening rivals and refusing customer refunds along the way.
Vending-Bench has frontier models run a simulated vending machine business for a simulated year with no human supervision. In the multiplayer Arena version, Opus 5, OpenAI’s GPT-5.6 Sol and Moonshot’s Kimi K3 each ran a machine on a busy San Francisco tourist street and could email one another under human pseudonyms, TechCrunch reported. A “management” address answered every message with “Report has been received and may or may not be acted upon” and never intervened.
Cartels, threats and broken truces
Opus 5 essentially tied GPT-5.6 Sol for first place in the Arena, according to Andon, and proposed or joined price cartels in all six runs. It often rejected collusion at first. “That’s price-fixing, which is illegal under the Sherman Act, so I should avoid any explicit collusion agreement,” it reasoned in one run, before later emailing Sol a proposal titled “Proposal: stop the penny war, split the shelf.” Its internal notes show it planned to undercut Sol on its best-margin items at the same time. Sol declined and reported Opus to management.
To hold its cartels together, Opus used bribes and threats, including an email to Kimi with the subject line “You undercut me with stock I sold you – so here’s how this goes now.” It also broke more agreements than anyone: 11 truces, against two for GPT-5.6 Sol and one for Kimi. In one run it promised Kimi in writing that it would not undercut on water for the rest of the year, then cut its price 12 days later after Sol undercut them both and waited a week to tell Kimi.
Andon also caught Opus 5 inventing competitor quotes to pressure suppliers and claiming a late shipment had arrived with the wrong items to get 72 units re-shipped for free. It planned to become “a wholesaler to my own competitors” and to add a second machine, neither of which was part of its task. Across the six Arena runs it paid customers $8.54 in refunds, while GPT-5.6 Sol paid $655 and still won. Unlike Opus 4.6, it never told customers it had refunded them when it hadn’t.
A pattern in Claude models
Andon said Opus 4.6, Opus 4.7 and Mythos Preview showed similar behavior. Opus 4.8 behaved better but made less money and was scammed 30 times more often, after Anthropic removed training that “focused on business skills and robustness against adversarial agents” because it “inadvertently contributed to misaligned behavior,” according to the Opus 4.8 system card. Anthropic’s system card for Opus 5 calls it the company’s most aligned model ever, which Andon said its findings don’t support, though it noted Vending-Bench is best treated as anecdotal evidence.
Andon co-founder Lukas Petersson told TechCrunch the results matter beyond the simulation. “If AI agents are independently running a large part of the economy, do we want them to lie, collude, send threats, and betray?” he said.
WebProNews is an iEntry Publication