Better Models: Worse Tools

This article by Armin Ronacher explores a surprising issue where newer, more advanced Claude models (Opus 4.8, Sonnet 5) produce worse tool call accuracy than older models. The author details how these models invent extra, non-schema fields in nested edits[] arrays when calling Pi's edit tool, leading to validation failures. The article explains LLM tool call internals, how models serialize function calls, and why this regression in SOTA models is concerning. It delves into technical aspects of AI model behavior, tool schema adherence, and the implications for software development tools relying on LLM integrations.