Anthropic engineer says Claude's writing declined as newer models focus on code
Jackson Kernion, an Anthropic engineer who works on Claude fine‑tuning, said the model’s writing quality has stalled or even worsened despite improvements in math, code, and reasoning. He explained that after Opus 4.6 – which he called the last good "writing model" – later versions were optimized for technical tasks and trained to produce explanations aimed at other AI models. This shift created…
Key points
- Jackson Kernion says newer Claude models prioritize math and code over human‑readable writing.
- Opus 4.6 was the last Anthropic model praised for writing quality, according to Kernion.
- Opus 5.5 improves the balance but has not been proven to beat Opus 4.6 in writing.
Kernion said the change stems from the reinforcement‑learning reward structure: rewards that favor AI‑model comprehension grow as models are trained more on code and math, while rewards for human‑friendly explanations receive less emphasis. With Opus 5.5, Anthropic has tried to rebalance the rewards, and Kernion noted he has not been "as happy about a model's writing since Opus 4.6." He cautioned that Opus 5.5 does not yet definitively surpass the older model in writing quality, and the problem remains an ongoing challenge.
Model page: Opus 5.5 →
The story so far
2 episodes →- Anthropic engineer says Claude's writing declined as newer models focus on codethis story
Anthropic engineer explains why Claude's writing got worse although the model got smarter
The Decoder · 23 September 2026
Loading the full article…
This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sourcesThe headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Liquid AI releases d1-ultra decision model at $0.042 per million tokens · 1 src
- Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal model with 42B active parameters and 1M-token context · 21 src
- Google launches Gemini 4 Argon model globally for Pro and Ultra subscribers with introductory API pricing · 32 src
- Microsoft AI chief says in-house model beats OpenAI's using quarter of compute · 1 src
- NVIDIA releases Nemotron 3.5 Super VL and Bolt edge models for enterprise agents · 8 src
Comments
via GitHub Discussions