Better Models: Worse Tools
TLDR
Newer Anthropic Claude models (Opus 4.8 and Sonnet 5) sometimes fail to properly call external tools by inventing extra fields in the API schema, whereas older models did not exhibit this problem. The failure rate reaches approximately 20% in some agentic contexts and can be reduced to zero by enabling strict tool invocation mode. This regression appears to result from post-training on Claude Code's forgiving tool harness, which silently repairs malformed calls and accepts parameter aliases, causing newer models to learn looser compliance with unfamiliar tool schemas.
Why it matters
Newer models get better at solving tasks while getting worse at faithfully emitting alternative tool schemas, requiring stronger guarantees in the harness.
