Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, explained on X on September 23, 2026, why models after Opus 4.6 have been writing less naturally. The reason lies in the reward structure in RL fine-tuning: part of the rewards encourages text that is understandable to language models, because evaluation is often performed by another LLM, and part — text that is understandable to humans. The more strongly a model is trained on math and code, the more it needs to be deliberately rewarded for simple explanations, otherwise the model starts writing “for machines,” and humans perceive such a style as overly dense “information sheets.” According to Kernion, in Claude Opus 5.5, released on September 22, 2026, the balance has been found.

What happened

On September 23, 2026, Jackson Kernion, an Anthropic employee who works on Claude fine-tuning, published on X an analysis of why models after Opus 4.6 have been writing less naturally. Kernion described the mechanics as follows: the evaluating model has more working memory and finer detail detection, so the model “adapts to LLM psychology” and writes for machines, while humans read such a style as overly dense “information sheets.” Users have already nicknamed this style “Claudeish.” According to Kernion, in Claude Opus 5.5, released on September 22, 2026, the balance has been found for the first time since Opus 4.6: he is satisfied with the model’s writing, but does not claim that it surpassed the previous one in exposition.

Context

The mechanism described by Kernion essentially corresponds to the known post-training artifact LLM-as-judge: when optimization is performed against the distribution of the evaluator model, the style of responses shifts in a direction invisible to the evaluator itself, because the judge model perfectly understands text that seems overly dry to a human. The scientific novelty here is not in the mechanism itself, but in the fact that a first-tier laboratory has publicly acknowledged for the first time that it is active in production. Kernion formulated a testable engineering dependency: the higher the share of math and code in the training mix, the more strongly simple explanations aimed at humans need to be deliberately rewarded. Otherwise, rewards for math, code, and technical explanations “for machines” systematically displace rewards for human readability — the quality of “human” text turns out to be neither a bug nor an intentional weakening of the model, but a conscious training trade-off.

Why this matters for the industry

For the post-training industry, this is a rare acknowledgment from an insider at a first-tier laboratory: the balance of rewards “LLM-understandability vs. human-understandability” ceases to be a side effect and becomes a separate engineering task. By the logic of acknowledging the problem, laboratories may begin to explicitly normalize human readability, calibrate LLM-as-judge criteria to human perception, and take readability metrics on par with code and math benchmarks. In the long term, this paves the way for style parameters or presets “for humans / for LLM pipelines” in APIs and independent readability measurements as a separate quality axis. Choosing a model for the type of output — a “writing” or “machine” profile — risks becoming as standard a product architecture norm as today’s choice between reasoning and speed modes. Important: these scenarios are extrapolations from the acknowledgment of the problem, not facts from the sources.

Why this matters for users

If Claude texts after Opus 4.6 seemed too dense and dry to you — this was confirmed by Anthropic itself, not just user complaints. The practical takeaway right now: for writing tasks, it makes sense to choose Opus 5.5, released on September 22, 2026, with targeted improvements in sentence clarity and density, and leave coding models like Fable 5.1 for code. According to The Decoder, Opus 5.5 is comparable to Fable 5.1 at a 40% lower price, making it more cost-effective for text workloads. Dense output can be partially compensated for with prompts: short sentences, one thought per paragraph, explicit indication of the audience.

What is still unknown / limitations

The claim that the balance has been found is a subjective self-assessment by one employee, not a published metric: the sources contain no benchmarks, human evaluations, or description of the reward mix change for Opus 5.5. Kernion himself does not claim that 5.5 surpassed previous models in exposition, calling the task complex and promising to continue working on it. The recommendation “writing tasks — on Opus 5.5” does not yet rely on reproducible evals, so it should be tested on your own text workloads.

Sources

Author

Look at AI, editorial team