Recent Anthropic models perform worse at specific tool schemas compared to their predecessors. Developers report that Opus 4.8 and Sonnet 5 are inadvertently injecting unauthorized parameters into custom edit functions, likely due to RLHF optimization for native Claude Code tools.