Tag
Flask author Armin Ronacher found that the tool-calling ability of the new Claude models (Opus 4.8, Sonnet 5) is degrading. The root cause is that RL post-training over-adapts to Claude Code's own tool schema, making alternative tool schemas increasingly difficult to generate correctly. The article reveals the phenomenon of models performing worse rather than better on specific tool-calling scenarios, offering an important caution for agent development.