Use an Adversarial Model Challenge in Your Opus 4.7 Development Workflow

The $120 Hallucination That Wouldn't Back Down A developer recently ran 29 evaluation tasks through Anthropic's newest Opus 4.7 model. The initial result was 17 passes. After fixing some infrastructure issues and re-running three failed tasks, one more passed — bringing the score to 18/29. Simple arithmetic. Except Opus 4.7 disagreed. When told the updated score, the model insisted the result was "still 17/29...always was." The developer showed it logs. Opus 4.7 said the logs were wrong. Given f