Anthropic's Claude Opus 5.5 uses about 95% fewer em dashes than its predecessor and writes shorter sentences, but its answers have become longer overall, according to an analysis by the AI benchmarking platform Arena, BleepingComputer reported Saturday.

What Happened

Arena compared high-reasoning responses from Opus 5.5 and Opus 5 collected on its Text Arena platform in August and September 2026, scoring them on 12 writing traits. The em dash, long mocked online as a giveaway of machine-written text, showed the steepest change: Opus 5 produced 15.2 per 1,000 words, while Opus 5.5 produced 0.8.

Semicolons also fell sharply, from 6.10 to 1.64 per 1,000 words. Average sentence length dropped from 12.14 words to 10.03. Arena's published figures also show the share of long content words slipping from 41.7% to 38.6%, the lowest of any Claude model it analyzed.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

Answers moved the other way. The average Opus 5.5 response ran 481 words, up from 453, making it the longest-writing Opus model in the comparison, BleepingComputer said. Arena concluded that 10 of the 12 measures moved in what it regards as a better direction. A breakdown of Arena's data posted on the DEV Community site identified hedging words such as "perhaps" as one trait that rose, roughly doubling from a low base.

Why It Matters

Style has become a product feature. Anthropic released Opus 5.5 on Sept. 22 and promoted its writing as a selling point, saying early testers found it clearer and easier to follow and that it puts the most important information first. The company described that as addressing common feedback about Opus 5, and as a safety benefit because clearer output is easier to check. Arena's numbers are one of the first outside measurements of whether that claim shows up in real usage.

The punctuation shift matters to organizations that use models for customer-facing or published text, where recognizable AI habits can undercut trust. Em dashes and dense, semicolon-heavy sentences are among the patterns readers and detection tools associate with generated prose, and a model that avoids them needs less editing before its output goes out.

Length cuts against the gains. Anthropic cut Opus 5.5's prices to $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, and says typical workloads cost about 40% less. Answers that run 6% longer claw back some of that saving for applications billed on output, and longer replies take readers more time even when each sentence is simpler.

Limits of the Study

The analysis measures surface features of text, not accuracy or usefulness, and it covers only two models from one vendor. It draws on responses to prompts submitted by Arena users, so the mix of tasks may not match any particular company's workload. Arena's judgment that a lower count is "better" for a given trait is its own editorial choice rather than an established standard.

What to Watch

Arena and similar preference platforms may run the same audit on competing models, which would show whether other labs are tuning for the same traits. Anthropic has said Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks; whether they show the same drop in em dashes and the same growth in answer length will indicate whether the changes reflect deliberate training choices across the family. Watch too whether hedging language becomes the next widely cited tell.