Why Claude Started Writing Drier Text Even as It Got Smarter

Jackson Kernion is an Anthropic engineer who works on fine-tuning Claude models. He explained why newer versions write drier - even though they clearly got smarter. The reason is models got tuned mostly for code and math. Explanations got polished for other AI systems to read, not for people (this is how the style grew that Kernion himself compares to jargon of a closed professional group). He said he wasn't happy with this kind of writing since Opus 4.6. Only with Opus 5.5 he found a better balance, but he stressed there's no real superiority over the old model here.

In my own use of Claude Code this was really noticeable. Detailed technical explanations read like a report, not like a conversation, and that got in the way when I needed text for a person (not code). For generating code itself this density of style doesn't hurt. But in comments, in writing to a client, or in text for a website it turns into a problem. If Opus 5.5 really brought the balance back closer to Opus 4.6, that's good news - not for writing code, but for tasks where the agent prepares text a real person will see.