🤖 Anthropic employee explains why Claude's writing got worse
Jackson Kernion, who works on fine-tuning Claude, explained on X on September 23, 2026: models since Opus 4.6 write less naturally due to the reward structure in RL fine-tuning. Some rewards favor text that is understandable to language models (evaluation is often performed by another LLM), while others favor text understandable to humans. The more a model is trained on math and code, the more simple explanations need to be rewarded, otherwise it 'adapts to LLM psychology': the evaluator model has more working memory and finer detail detection, so a style convenient for the model is read by humans as overly dense 'information sheets' — users have called it 'Claudeish'. In Claude Opus 5.5, released on September 22, Kernion says the balance has been found — he is satisfied with the model's writing for the first time since Opus 4.6, but does not claim that the new version has surpassed the old one: 'the task is complex, we will continue working on it'.
🌍 This is a rare public admission from an insider at a top-tier lab: the quality of 'human' text is not a bug or a 'nerf,' but a conscious trade-off in RL training. Rewards for math, code, and technical explanations 'for machines' systematically crowd out rewards for readability, and the balance of 'LLM-understandability vs human-understandability' becomes a separate engineering task in post-training for the entire industry.
👤 If Claude's texts since Opus 4.6 seemed too dense and dry to you — this was confirmed by Anthropic itself, not just user complaints. For writing tasks, it is now more reasonable to try Opus 5.5, and leave coding models like Fable 5.1 for code; note that 'satisfied with the model's writing' is the self-assessment of one employee, not a published metric. Dense output can be partially compensated for with prompts: short sentences, one thought per paragraph, explicit indication of the audience.
Source 1: https://x.com/JacksonKernion/status/2102437423670325581 Source 2: https://the-decoder.com/anthropic-engineer-explains-why-claudes-writing-got-worse-although-the-model-got-smarter/