Anthropic benchmark pushes task-based evals over leaderboards

A third-party breakdown claims Anthropic introduced a new benchmark alongside recent Claude updates, emphasizing process-based, tool-using reasoning instead of static leaderboard scores. For engineering teams, the takeaway is to evaluate LLMs on end-to-end tasks (retrieval, code/SQL generation, execution, and verification) rather than rely on single-number accuracy.