Better Models: Worse Tools

According to developer Armin Ronacher, newer Anthropic Claude models like Opus four point eight and Sonnet five have started making errors when using custom edit tools. The models are inventing extra fields that don't match the tool's schema, causing calls to fail. Surprisingly, this is getting *worse* with newer models—older Claude versions handle it fine. Ronacher theorizes this is a side effect of training newer models to excel at Claude Code's specific edit tool. The irony: better models for one coding harness may be worse for others. It raises a question for third-party tools: should they implement multiple edit schemas just to match whichever model their users choose?

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton