Arena Analysis: Opus 5.5 Cuts Em Dashes 95% and Reads Less Like AI, but Answers 6% Longer
rohanpaul_ai · x · 2026-09-27
- Arena analyzed Claude Opus 5.5's writing against Opus 5 on high-reasoning Text Arena outputs: 10 of 12 writing measures improved.
- Fewer AI tells: em dashes fell from 15.2 to 0.8 per 1,000 words (a 95% drop) and semicolons fell 73%; long content words dropped from 41.7% to 38.6%, the lowest of any Claude model analyzed.
- More readable but longer: average sentence length fell 17% (12.14 to 10.03 words), yet answers grew 6% wordier (453 to 481 words), the longest in the Opus family.
- A new tell may be emerging: increased hedging and caveats.
Related event: Claude Opus 5.5 Writes More Human-Like, Em Dashes Down 95%(4 posts)→
More from Models
- Weekly: Amazon blocks Muse, agents fight over restaurant bookings, and Opus vs GPT price war — njyx · 2026-09-27
- Blind test of 90 page pairs: named style references beat adjectives 68% across Opus, Sol, Kimi — maxsloef · 2026-09-27
- Eval design note: models self-select style references in a separate call to isolate the effect — maxsloef · 2026-09-27
- LLM graders can't match human taste on style eval — maybe why the gap persists — maxsloef · 2026-09-27
- Namedrop eval: style references beat adjective prompts even when models pick the reference — maxsloef · 2026-09-27
- Interactive Architecture Atlas Maps DeepSeek V4, GLM 5.3 and Kimi K3 Down to Kernel Tensor Shapes — zhyncs42 · 2026-09-27