Claude Opus 4, Claude Sonnet 4
Anthropic
partial detail 15 method fields published
Evaluation
Company-wide model testing exercise (pre-launch)
issues reported by employees testing both Claude 4 models in roughly final forms (mild harmlessness, sycophancy, hallucination, reward-hacking-related behavior)
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.