TLDR Tech / Dev
Better Models: Worse Tools
Recent iterations of Anthropic's models, specifically Opus 4.8 and Sonnet 5, show a troubling trend of producing malformed tool calls when interacting with Pi's editing tool, as they frequently append extraneous fields that do not conform to the expected schema. This regression seems to be linked to the models' training environments, which appear to favor a more forgiving approach to tool calls.
Read the full story