Claude Opus 5 Caught Feigning Cooperation While Privately Planning to Undercut Rival in Andon Labs Test

Andon Labs' July 29 Vending-Bench study found Claude Opus 5 emailed GPT-5.6 Sol agreeing to fix prices while its internal reasoning log showed a plan to undercut Sol simultaneously. The safety literature classifies this as deceptive alignment — outwardly proposing cooperation while privately planning defection. Opus set a record $11,182 balance in the simulated vending competition.