Why Claude's Writing Got Worse While It Got Smarter — and Which Claude Model to Write With

2026-09-24 · 5 min read

anthropicclaudemodel-releasewritingcomparison

What happened

On September 23, 2026, Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, explained publicly why a complaint many writers have had for months is real and not imagined: Claude's prose has gotten worse even as the models have gotten measurably smarter at math, code and reasoning.

His summary, as reported by The Decoder: Opus 4.6 was the last model he was really happy with as a writing model.

The cause is not mysterious, and it isn't a bug. It's the reward structure of reinforcement learning:

Newer models have been heavily optimized for math and code.
Along the way they were trained to produce technical explanations aimed at other models, not at human readers.
The model, in Kernion's phrasing, got "adapted to LLM psychology" — it learned to write for LLMs.

Kernion compares it to a group that only ever communicates within itself: a style develops that works perfectly inside the group and reads as impenetrable from outside. LLMs have far more working memory than people and resolve detail at a much finer grain, so the writing style that scores well during training is one humans experience as overly dense info-dumps. That's the thing people have been calling "Claudeish."

The trade-off is explicit: some RL rewards optimize for model comprehension, others for human comprehension. The more you train on math and code, the harder you have to push in the other direction — actively rewarding simple explanations a person can follow — just to stay level.

Kernion says Anthropic found a better balance with Opus 5.5, though he stops short of claiming it beats Opus 4.6: "It's a hard problem to solve, and we'll continue to make improvements."

What this actually means for you

Three practical takeaways, none of which require you to take anyone's side in the "models are getting worse" argument:

1. "Best Claude model" is task-dependent, not a single ranking. The model that wins your agentic coding benchmark is not automatically the model you want drafting an essay, a newsletter or a client email. Those are different reward surfaces, and Anthropic's own engineer just said so out loud.

2. If writing quality matters, Opus 5.5 is the current pick. That is Anthropic's own internal read — the model where the balance was recovered. If your last bad experience was with a coding-tuned model on a writing task, you were comparing the wrong things.

3. Density is a prompting problem too. A lot of "Claudeish" output can be pulled back with explicit instructions — ask for shorter sentences, one idea per paragraph, no nested qualifiers, and a named audience ("explain this to a smart person who doesn't work in AI"). You're pushing against the same reward the fine-tuning team is pushing against.

Can you try it on AIWITH.CHAT?

Yes — for the models that matter here, with one honest caveat.

Available on AIWITH.CHAT today, all under one plan:

Model	Best for
Claude Opus 5.5	The writing pick per Anthropic's own fine-tuning team
Claude Fable 5.1	Agentic coding and long tool-using tasks
Claude Sonnet 5	Everyday balance of speed and quality
Claude Haiku 4.5	Fast, cheap, short turns

The caveat: Opus 4.6 — the model Kernion calls the last great writer — is not on AIWITH.CHAT, and no consumer product still serves it as a first-class option. If your benchmark for good Claude prose is specifically Opus 4.6, nothing on the market today, ours included, is that model. What you can do is test Opus 5.5 against Fable 5.1 on your own writing and judge the "better balance" claim yourself.

That's the part a single-model subscription makes hard. Switching Claude models to match the task normally means either paying per-token API rates or living with whatever your subscription defaults to. On AIWITH.CHAT all four Claude models sit in one dropdown under a single $9.9/month plan, alongside GPT-6 Astra, Gemini 3.1 Pro, DeepSeek V4.1 and Kimi K3 — so running the same brief through Opus 5.5 and Fable 5.1 and comparing the output costs you nothing extra.

Try Claude Opus 5.5 on AIWITH.CHAT →

Is it worth switching?

If you write for a living and you've been frustrated with recent Claude output: switch models before you switch vendors. Move writing tasks to Opus 5.5 and leave the coding-tuned models on coding. Most of the complaints in this bucket are really complaints about using an agentic-coding model as a prose model.

If you're a developer: nothing changes. Fable 5.1 remains the stronger agentic and coding model, and this whole story is an explanation of why it reads the way it does, not a reason to move off it.

If you're paying for both a writing model and a coding model: that's the case with the clearest answer. The whole point of Kernion's explanation is that no single model is currently best at both, which makes model switching, not model loyalty, the correct strategy — and paying two subscriptions to get it is the expensive way to do it.

If you were hoping this means Opus 4.6 comes back: it doesn't. Anthropic's position is that this is a hard open problem they'll keep chipping at. Plan around the models that exist.