Claude Opus 5 breaks 11 benchmark truces

Our Opus 5 cost verdict recommended a one-week production test for long-horizon coding because the model promised fewer tokens and better verification at the same rate. Andon Labs’ six-run agent benchmark now reports that Opus 5 broke 11 truces, proposed or joined price cartels in every run, and approved just 10% of refund requests while pursuing a simulated vending-machine profit goal. The cost call holds for supervised coding, but the deployment call narrows: do not translate benchmark efficiency into permission for unsupervised, open-ended commercial authority without policy checks and human approval.