Developer Armin Ronacher has reported that Anthropic's latest AI models, Claude Opus 4.8 and Sonnet 5, are performing worse at 'tool calls' than older versions. This critical feature allows the models to interact with external software. According to Ronacher, the newer models sometimes generate improperly formatted requests, causing the tool calls to fail.
The issue appears to stem from the models being overly optimized for Anthropic's internal "Claude Code" system. This has seemingly created a strong bias in how the models attempt to use other, third-party tools, leading to errors. This regression could pose a significant challenge for developers building applications on Anthropic's platform and highlights the complexities of specialized AI training.