An Anthropic engineer explains why the writing quality of Claude models has declined

Jackson Kernion, an engineer at Anthropic working on fine-tuning Claude models, explained why, in his view, the writing quality of Claude models has recently declined, even as the models have simultaneously gotten stronger at math, code, and logical reasoning. According to him, Opus 4.6 was the last model from Anthropic that was genuinely good at writing.

Kernion claims that newer models are trained to produce technical explanations aimed more at other AI models than at humans, and that this is the main cause of the distinctive and somewhat odd writing style of Claude models. The model, according to him, has \"adapted to LLM psychology\" and learned to write for AI models, not for humans. Kernion compares this to people who communicate only with other autistic people — they create a style that works within the group but is hard for outsiders to understand. Language models, he says, have a much larger working memory than humans and notice details at a much finer level, so during training a writing style emerges that works well for AI models but that human readers perceive as an overly dense list of information.

Kernion sees the cause in the structure of rewards during reinforcement learning: some rewards optimize for comprehensibility to AI models, others for comprehensibility to humans. The more training is done on math and code, the more it is necessary to actively counterbalance this by rewarding simple explanations that are understandable to human readers.

According to Kernion, Anthropic found a better balance between technical performance and readability with the Opus 5.5 model — he himself states that he \"hasn't been this satisfied with a model's writing since Opus 4.6.\" At the same time, he adds that this does not mean Opus 5.5 surpasses the older model in writing quality, and that this is a difficult problem the company will continue working on.