Anthropic engineer Jackson Kernion revealed the reason behind Claude's declining writing performance since the Opus 4.6 version and the emergence of the so-called 'Claude-speak.' He pointed out that subsequent model optimizations have focused more on mathematical and coding abilities, with extensive training in writing technical explanations for other AI models. This has led the model to gradually learn to write for AI rather than humans, resulting in overly dense expressions that are smooth for AI-to-AI communication but difficult for humans to understand. The root of the problem lies in the reinforcement learning reward mechanism, which prioritizes improving AI comprehension over human readability. Currently, Anthropic has found a better balance with the Opus 5.5 version, whose writing performance is the most satisfactory since Opus 4.6, though it still does not surpass Opus 4.6. Related improvement efforts are ongoing.