Anthropic Engineer explains Claude Writing Change: Modeling is written to AI, not to people

Anthropic Engineer Jackson Kneen explained the decline in Claude's writing performance from Opus 4.6, which formed the so-called Claude cavity: the follow-on model optimized the focus on mathematical and code capabilities, and was heavily trained in other AI model writing techniques, which learned to write to AI rather than to humans, creating an over-intensity expression of AI that is smooth but difficult for humans to understand, rooted in an enhanced learning incentive mechanism that focuses on upgrading AI understanding effects rather than human readability. At present, Anthropic has found a better balance on Opus 5.5, whose writing performance is the most satisfactory since Opus 4.6, but has not yet surpassed Opus 4.6, and improvements continue。