Evaluation

Reward hacking evaluations (reward-hack-prone coding tasks, Claude Code impossible tasks)

reward hacking rates across three settings: reward-hack-prone coding tasks (classifier scores and hidden tests), Claude Code impossible tasks (unsolvable tasks with anti-hack prompt variants), and training-distribution tracking

Model behavior & alignment Anthropic

Ask about this evaluation

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.

Claude Opus 4, Claude Sonnet 4 May 2025

Anthropic

rich detail 30 method fields published

See how this test was run