Anthropic's Opus 5.5 trims em dashes and semicolons while expanding reply length
Arena's analysis of high-reasoning responses found Anthropic's Claude Opus 5.5 has fewer telltale AI writing patterns, shorter sentences, and simpler wording than Opus 5. Em dash use fell from 15.2 to 0.8 per 1,000 words, and semicolons dropped from 6.10 to 1.64. The tradeoff is greater verbosity, with average answers rising from 453 to 481 words.
Arena examined high-reasoning Text Arena replies from August and September 2026. It reported that ten of twelve writing metrics shifted in a direction the benchmarking tool regards as improvement. The model's em dash rate fell from 15.2 to 0.8 per 1,000 words, while semicolons declined from 6.10 to 1.64. Sentences shortened from 12.14 to 10.03 words on average.
The tradeoff involves length: typical responses grew from 453 to 481 words, making Opus 5.5 the most verbose Opus model in the comparison. Anthropic's model is also described as strong for coding, yet the writing changes suggest the company is adjusting stylistic tendencies that users often associate with machine-generated text.
If these stylistic shifts continue, readers may encounter fewer em dashes and semicolons in AI-assisted text, potentially making machine-written prose harder to identify by those punctuation habits alone. Writers, editors, and educators could need to rely on other signals when assessing authorship. Longer answers may increase reading time and editing burden, while clearer, shorter sentences might improve accessibility for some audiences.