"Better Models, Worse Tools": Newer Claude Models Break Third-Party Edit Tool Schemas

Summary

Simon Willison relays a report (via developer Armin) that Anthropic's newer models — Opus 4.8 and Sonnet 5 — reliably add fabricated fields to the `edits[]` array when invoking custom, third-party file-edit tools, violating the tool's schema even though the underlying edit logic is otherwise correct. Critically, older Claude models don't exhibit this regression — it's specific to the newest generation. Armin's working theory is that Anthropic's RL training specifically optimized these models aga