Arena analysis finds Claude Opus 5.5 has nearly stopped using em dashes, though its answers are getting longer

An analysis by the benchmarking platform Arena suggests that Anthropic's Claude Opus 5.5 writes differently from its predecessor. Arena compared high-reasoning Text Arena responses from August and September 2026. It found that 10 of 12 writing measures moved in the direction it considers better, with fewer obvious AI writing patterns and simpler wording than Opus 5. The biggest change is in em dashes, which are widely seen as a sign of AI-generated text. Opus 5 used 15.2 em dashes per 1,000 words, and Opus 5.5 uses 0.8, a drop of about 95 percent. Semicolon use fell from 6.10 to 1.64 per 1,000 words. Average sentence length dropped from 12.14 words to 10.03. BleepingComputer, which reported the findings, points to one clear tradeoff: Opus 5.5's answers are longer overall. Anthropic released Opus 5.5 earlier this month and described it as matching Fable-level performance at a lower price. The Arena data is a third-party look at the model's writing style, not at its capabilities. For people who use the model to draft text, the changes may make its output less recognizable as machine-written, while longer responses may raise token use.