
Claude Opus 5.5 slashes em dash usage by 95 percent
Anthropic's Claude Opus 5.5 uses 95% fewer em dashes, writes shorter sentences, yet its answers grow longer, according to new benchmark analysis.
Anthropic's Claude Opus 5.5 is changing how it writes, and one of the most visible signs is a near-total retreat from the em dash. According to fresh measurements published by Arena, an AI benchmarking platform, the latest model uses roughly 95 percent fewer em dashes than its predecessor, Opus 5. That is a drop from 15.2 em dashes per 1,000 words to just 0.8 per 1,000 words in responses collected during August and September 2026. The semicolon is also losing ground; its frequency fell from 6.10 to 1.64 per 1,000 words over the same period. Because heavy em dash usage has long been one of the mechanical fingerprints of AI-generated prose, this shift could make the model's outputs harder to spot as machine-written.
The writing overhaul runs deeper than punctuation. Arena reports that sentences have become shorter, averaging 10.03 words compared with 12.14 words in Opus 5. The wording itself is simpler and less convoluted. In total, 10 out of 12 writing metrics that Arena tracks moved in what it considers a better direction, meaning fewer repetitive structures, less jargon, and a more natural rhythm. This suggests Anthropic is deliberately tuning the model away from the stiff, over-punctuated style that often marks content as "AI slop" (the industry term for low-quality, generic machine-generated text).
There is one clear trade-off. While individual sentences are shorter, the complete answers are growing longer. The average Opus 5 response ran 453 words; Opus 5.5 pushes that to 481 words, making it the most verbose version of Claude in the comparison set. A longer reply does not automatically mean a better one, but for users who rely on the model to draft articles or blog posts, the extra length may deliver more substance or, conversely, more fluff. The real impact depends on how site owners and editors rework that raw material before hitting publish.
Beyond the numbers, the shift matters because most recent AI model announcements have focused on coding ability, not on human-like writing. Anthropic's decision to refine Claude's prose suggests the company believes there is still substantial room to improve how AI communicates. For website owners, marketing teams, and small businesses that use these tools to produce first drafts of web copy, fewer obvious mechanical habits mean less time spent scrubbing the telltale signs of machine authorship. However, a smoother draft is not the same as a fact-checked or original one, so human oversight remains essential.
For anyone managing a website where AI-assisted content appears, the lesson is that polish alone does not build trust. Visitors care about speed, uptime, and security just as much as they care about what the words say. A thoughtfully written page loses its impact if the site loads slowly or goes offline. Choosing a hosting platform that prioritises performance and protection, such as AEU Hosting, ensures that the foundation beneath every published paragraph stays solid, no matter which writing tool helped compose it.
How to Protect Yourself
- Always verify any specific details, like statistics or product names, that AI tools produce for your content.
- Read AI-generated drafts aloud to spot unnatural phrasing and smooth out sentences before publishing.
- Keep your website's hosting core up to date and protect it with strong passwords and regular backups, regardless of how you create your pages.
Terms Explained
- em dash A long dash used in writing to separate thoughts; many AI models tend to overuse it, which helps readers spot machine-generated text.
- semicolon A punctuation mark used to connect closely related independent clauses, often overused in stiff AI writing.
- AI slop Low-quality, generic text produced by artificial intelligence without thoughtful human editing or clear purpose.
- verbose Using more words than necessary to express an idea, making the message longer but not always clearer.
- benchmarking The process of measuring a model's performance or output against a set of standards to compare it with other models.