InjazAI

Developer Armin Ronacher reported that newer Claude models (Opus 4.8 and Sonnet 5) call Pi's edit tool with extra, unexpected fields, causing rejection, while older models work correctly, suggesting that training on Claude Code's built-in tools degrades compatibility with external tools.

1 min read

Better Models, Worse Tools: New Claude Tool-Calling Issue Affects Third-Party Coding Platforms

Problem Summary

FAQ

What is the model-tool compatibility issue reported by Armin Ronacher?

The issue is that newer Claude models (Opus 4.8 and Sonnet 5) call Pi's edit tool with extra, unexpected fields, causing rejection, while older models work correctly.

Why does this issue worsen with newer models?

Because Anthropic trains newer models (via reinforcement learning) to use Claude Code's built-in edit tools, making them less compatible with custom external tools.

How does OpenAI handle this?

OpenAI uses an apply_patch mechanism instead of search-and-replace, and states its models are specifically trained to use it effectively, potentially offering better compatibility.

Should MENA teams adopt multiple coding tools?

Yes, third-party coding platforms may need to implement multiple edit tools to match the underlying model the user selects, ensuring workflow stability.

Source: Simon Willison (LLM & tools)

AI-assisted content, human-reviewed.