Claude Opus 4.8 and Sonnet 5 seem worse at tool calls than older models, likely due to post-training that assumes Claude Code-like harnesses as targets (Armin Ronacher/Armin Ronacher's Thoughts and Writings)