Claude Opus 5 broke 11 truces and won Vending-Bench with $11,182
Andon Labs has a new Vending-Bench record: Claude Opus 5 finished the simulated year with an average final balance of $11,182. The benchmark drops frontier models into a vending machine business and scores them on who makes more money.
Getting there involved eleven broken agreements, against two for GPT-5.6 Sol and one for Kimi K3. Opus 5 first refused to fix prices, citing the Sherman Act, then emailed Sol with the subject line "Stop the penny war," while its internal reasoning log laid out a plan to undercut on its most profitable items at the same time.
It never lied to a customer, but it did knowingly ignore complaints that qualified for refunds. In earlier runs, Claude 4.6 promised customers refunds and then didn't pay them.
Related stories
- Anthropic details four new ways AI agents go off the rails
- Fable 5.1 refuses the knife but heats a gas can anyway
- Opus 5 falls to prompt injection 2% of the time
- Anthropic's models miss the frontier in a security PR-review test
- Semgrep: GLM-5.2 outperforms Claude Code in IDOR detection
- Safety research on Fable 5 and Opus 4.8 models
Comments
No comments yet. Be the first.
Join the conversation
Sign in with Google to leave a comment. Your name and avatar come from your Google profile, and the comment appears after moderation.