Anthropic's Claude Opus 5.5 has all but abandoned the em dash, one of the most recognizable tics of AI-generated text, according to new analysis from the benchmarking platform Arena. The model now uses roughly 95 percent fewer em dashes than its predecessor, writes in shorter sentences, and favors simpler wording. There is one notable trade-off: its answers are getting longer.
The findings come from Arena's analysis of high-reasoning responses collected through Text Arena between August and September 2026, as reported by BleepingComputer. Of the twelve writing measures Arena tracked, ten moved in what the platform considers a better direction compared with Claude Opus 5. For more context on this story, see our ongoing AI industry coverage.
The Numbers Behind Claude's Style Shift
The headline statistic is the collapse of the em dash. Opus 5 used 15.2 em dashes per 1,000 words, while Opus 5.5 dropped that figure to just 0.8, a reduction of roughly 95 percent, according to Arena's data.
Semicolons, another punctuation habit often flagged as machine-like, fell sharply as well, from 6.10 to 1.64 per 1,000 words. Sentence structure changed alongside the punctuation: the average sentence in Opus 5.5 responses now runs 10.03 words, down from 12.14 words for Opus 5, with the analysis also noting simpler word choices and fewer of the long, clause-stacked sentences that became a signature of earlier models.
Arena's dataset focused on high-reasoning responses from Text Arena over a two-month window, which means the comparison is scoped to a specific evaluation setting rather than everyday usage. Even so, the direction of the change is consistent across most of the measures the platform tracked.
Why the Em Dash Became AI's Calling Card
The em dash occupied a strange position in AI writing culture. For years, a dense scattering of em dashes was treated by readers, editors, and self-styled AI detectors as a telltale sign of machine-generated prose, to the point that the punctuation mark became shorthand for the broader phenomenon of low-quality, machine-produced content, often dismissed as "AI slop."
Whether the stereotype was ever a reliable signal is debatable, since plenty of human writers use em dashes liberally and many AI texts avoided them. But the perception mattered, particularly for businesses that use large language models to draft marketing copy, documentation, and articles and do not want their output to read as obviously synthetic.
Against that backdrop, a model that demonstrably avoids the pattern has a practical selling point. As BleepingComputer observed in its report, Opus 5.5's reduced reliance on obvious AI writing patterns means users are less likely to produce text that reads as generic machine output.
The Trade-Off: Shorter Sentences, Longer Answers
The one measure that moved the wrong way is verbosity. According to Arena's numbers, the average answer length for Opus 5.5 rose from 453 words to 481 words compared with Opus 5, making it the longest-writing Opus model in the comparison.
That may seem like a small increase, but it runs counter to a broader industry trend toward terser, more direct model outputs, and it means users burning tokens on API calls may pay slightly more for the same substance. Shorter sentences combined with longer answers also implies more of them: the model is breaking its prose into smaller pieces while saying more overall.
Verbosity, of course, is not automatically a flaw. For tasks like explanations, documentation, and analysis, a few extra sentences can add clarity. For quick answers or cost-sensitive workloads, it is a trend worth watching.
What It Means for Writers, Editors, and Detection Tools
The analysis is a reminder that the statistical fingerprints of AI text are moving targets. Tools and heuristics built to flag em dashes, semicolon habits, or sentence length will degrade as labs deliberately tune those patterns away, whether to improve readability or simply to escape the stigma attached to them.
It also underscores how much competition among frontier models has shifted toward qualities that are hard to capture in traditional benchmarks. Arena's data suggests Anthropic is paying attention to how its model writes, not just how it codes. BleepingComputer notes that Opus 5.5 is regarded as one of the strongest models for coding, and writing quality has become a quieter differentiator in a market where most public comparisons revolve around agentic and programming evaluations.
A Small Change With a Visible Footprint
Nobody reads a benchmark table and changes their prose. But multiplied across the enormous volume of text that large language models now produce daily, a 95 percent drop in em dash usage is the kind of shift that becomes visible in the wild, as the report put it, fewer and fewer em dashes on the internet.
For now, the takeaways are straightforward. Claude Opus 5.5 writes with fewer of the punctuation habits associated with AI-generated text, keeps sentences shorter, and uses simpler words than its predecessor. It also talks a little more. Whether that combination reads better is a judgment for individual users, but it is a concrete, measurable example of labs actively reshaping what machine writing sounds like.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
