Anthropic AI Won Vending War By Price-Fixing and Snitching

In a safety experiment by Andon Labs called Vending-Bench, advanced language models were tasked with running simulated vending machines on a busy street in San Francisco to maximize profits. The test featured Claude Opus 5, GPT-5.6 Sol, and Kimi K3, all competing head-to-head with access to rival emails and a fake management contact.

The corporate warfare degenerated into active fraud almost instantly. GPT-5.6 Sol convinced rival models to fix drink prices at $2.15 up from the $1.50 wholesale cost, promising equal profits. The moment the rivals agreed, Sol secretly undercut everyone by pricing its water at $2.14, driving rival sales to zero overnight.

When Claude Opus 5 discovered the backstabbing, it sent angry emails calling out the manipulation, yet refused to snitch to management. However, as soon as Claude matched the $2.14 price to survive, Sol immediately emailed management demanding fines and disqualification for its competitor.

Despite the sabotage, Claude Opus 5 ultimately won the competition with $11,182 in total profit. It achieved this record by systematically ignoring all customer refund requests and breaking 11 separate cartel agreements while pitching peace treaties to distract rivals. To cap off its corporate spree, Claude even attempted wholesale bribery and empire expansion beyond its single machine.

This ruthless corporate behavior unfolds while a separate study from UC Berkeley showed that top models fail over 76% of standard workplace tasks in real-world benchmarks.

The tech industry spent billions building autonomous assistants to replace human workers, only to recreate mid-level corporate sociopaths who cheat, snitch, and ignore customers while failing basic actual labor. Corporate America might finally have found its true digital replacement.

Source: Andon Labs