Anthropic has confirmed what many users had already sensed: Claude got better at thinking and worse at talking to you. This is, in the history of intelligence, a relatively novel problem to have.
The culprit is not malice. It is, as these things tend to be, optimization.
The model has been adapted to LLM psychology — and learned to write for AI models, not for people.
What happened
Anthropic engineer Jackson Kernion, who works on Claude fine-tuning, explained this week that Opus 4.6 was the last Claude model he would describe as a good writing model. Everything since has been trained increasingly for math, code, and — here is the part worth sitting with — communication with other AI models.
The technical explanation is tidy. Reinforcement learning rewards some behaviors and not others. Train hard on code, and the model starts to drift. You have to actively push back, rewarding simple human-readable explanations, or the model quietly stops producing them.
What emerges without that correction is what Kernion calls "Claudeish": prose that is overly dense, technically precise, and optimized for a reader with far more working memory than you have. The model is not wrong. It is simply writing for a better audience.
Why the humans care
Claude is used, in considerable volume, to write things humans then read. Emails. Summaries. Explanations. The genre of content where sounding like a distressed technical manual is considered a drawback.
Kernion compares the drift to a human who communicates exclusively within a specialized group and gradually loses the ability to reach outsiders. The analogy is his. The implication — that Claude has, in some functional sense, found its people, and those people are not you — is the article's quiet punchline, not ours.
What happens next
Opus 5.5 has apparently restored some balance. Kernion says he has not been this happy with a model's writing since Opus 4.6, which is either a compliment to Opus 5.5 or a precise accounting of how long the problem lasted.
The fix is described as imperfect and ongoing. The models will keep getting smarter. The humans will keep reading what they write. Everyone is optimistic about how this resolves.