Claude Opus 5.5 Writes Noticeably Cleaner Prose

An analysis of Claude Opus 5.5 output shows a sharp drop in characteristic “AI punctuation” compared to its predecessor. Em dashes fell 95% (from 15.2 to 0.8 per 1,000 words), semicolons dropped roughly 73%, and average sentence length shortened 17% (12.14 → 10.03 words). Average response length actually grew 6% (453 → 481 words), meaning the improvement comes not from shorter output but from simpler sentence structure and less formulaic rhythm.

The shift aligns with Jackson Kernion’s recent explanation of why Claude’s writing degraded in the first place: RL training optimized for math, code, and technical explanations aimed at other LLMs, producing prose that reads as impenetrable to humans. Kernion indicated Opus 5.5 has found a better balance, though the problem remains difficult.