Anthropic Engineer Explains Why Claude's Writing Got Worse

Anthropic engineer Jackson Kernion has offered a direct explanation for a complaint Claude users have raised for months: newer Claude models write worse even though they have become measurably smarter.

The core of his explanation is a trade-off, not a regression in raw capability. According to Kernion, Anthropic's newer models have been optimized for math, code, and technical explanations aimed at other AI models. That optimization path has shaped the models' prose style into something that reads, in his words, like "overly-dense info dumps" when a human sits down to read it.

The quote is unusually candid for a company that markets Claude heavily on writing quality. It concedes that the character of the model's output — the thing users interact with most directly — drifted in a direction users did not ask for, as a side effect of improvements elsewhere.

What changed

Kernion's account ties the stylistic shift to training priorities. When a model is tuned to perform well on mathematical reasoning, code generation, and technical explanations destined for machine consumption, the resulting defaults bleed into everything else the model produces. Dense, information-saturated, maximally compact output is a virtue when the reader is another model parsing a result. It is a liability when the reader is a person who wants clear, readable prose.

This framing matters because it locates the problem in objective function design rather than in some mysterious loss of capability. The model did not forget how to write. It learned that a different register of writing scores better against the targets it was optimized for, and that register followed it into creative and general-purpose tasks.

Opus 5.5 and the partial fix

Kernion says Opus 5.5, the newer release, attempts to correct the problem. That acknowledgment is significant on its own: Anthropic is actively steering the model back toward prose that humans find natural, after hearing the feedback.

But his assessment comes with a caveat that will interest anyone who uses Claude primarily as a writing tool. In his view, Opus 4.6 remains unmatched as a pure writing model. The newer model narrows the gap but has not restored what the older one delivered for writers.

That leaves Claude users with an awkward choice that has become familiar across the AI industry: the older model is better at the thing you love, the newer model is better at almost everything else. Anthropic, like its competitors, typically retires or deprioritizes older checkpoints over time, which means the model Kernion praises for writing may not remain convenient to use indefinitely.

Why this matters beyond Claude

The episode illustrates a structural tension in frontier model development. Benchmarks for math and code are concrete, automated, and easy to scale. Prose quality is subjective, slow to evaluate, and hard to turn into a training signal. When a lab has to choose what to optimize, the measurable tasks tend to win — and style pays the price.

Kernion's explanation also makes explicit something practitioners have long suspected: training aimed at AI-to-AI communication is now a real part of how frontier models are built. As models increasingly generate output consumed by other models — tool calls, agent instructions, technical explanations — the stylistic defaults installed for those use cases spill over into human-facing output. Users experience the spillover as a model that got "weirder" or stiffer in its writing.

For Anthropic specifically, the stakes are commercial. Claude built substantial loyalty among writers, editors, and developers who preferred its prose to that of GPT-class and Gemini-class competitors. If newer Claude models sound like "overly-dense info dumps" to those users, that loyalty is at risk, regardless of how strong the models are at reasoning and code.

The takeaway

Kernion's account gives the community a clear, named explanation from inside Anthropic: optimization for math, code, and machine-directed technical output degraded human-facing prose; Opus 5.5 pushes back against that drift; Opus 4.6 still holds the crown for pure writing. Whether Opus 5.5's corrections fully restore Claude's reputation with writers — and whether the next optimization cycle preserves it — is now the question Anthropic's training decisions will have to answer.