Anthropic engineer explains why Claude's writing got worse although the model got smarter

Anthropic's models have become dramatically better at mathematics, programming and multi-step reasoning. The quality of their prose has not kept pace, and in some cases it has fallen. Jackson Kernion, an Anthropic researcher who works on Claude fine-tuning, has offered an explanation for that gap in a series of posts on X.

Optimized for code, at the expense of natural language

According to Kernion, Opus 4.6 was the last Anthropic model he would call a genuinely good "writing model". Part of the reason is predictable: newer releases were optimized for math and code. But there is a second, less visible factor — the models were also trained to produce technical explanations aimed at other AI models rather than at people.

The model has been "adapted to LLM psychology", Kernion says, and learned to write for machines. That, he argues, is where the odd formulations users call "Claudeish" come from: text that looks precise and information-dense but is hard for a human to read.

An audience of language models

Kernion compares the effect to people who communicate mainly within one narrow group: they develop a style that works well inside it but is difficult for outsiders to follow. LLMs have far more working memory than humans and register details at a much finer level, so a style emerges during training that suits other models and lands with human readers as "overly-dense info dumps".

The cause lies in the reward structure

For Kernion, the root of the problem is the reward structure of reinforcement learning. Some rewards optimize for how well another AI model understands an answer, others for how well a human does. Heavy training on math and code shifts the balance toward the first kind, and the more a model is trained that way, the more developers have to push back by rewarding simple explanations a person can actually follow.

Anthropic found a better balance with Opus 5.5, Kernion says, adding that he has not been "as happy about a model's writing since Opus 4.6". He does not claim the newer model surpasses the older one. "It's a hard problem to solve, and we'll continue to make improvements," he writes.