From 15.2 to 0.8 Em Dashes per 1,000 Words: Claude Finally Stops Writing Like a Bot

The em dash became the internet's favorite AI litmus test. Spot one too many in a LinkedIn post? Bot. See three in a paragraph? Definitely bot. Well, the numbers are in: Claude 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%. The meme is officially dead.

Arena's writing analysis of Claude Opus 5.5

The data comes from Arena, an AI benchmarking tool that analyzed high-reasoning Text Arena responses from August and September 2026 and observed that 10 of 12 writing measures moved in what it considers a better direction. Sentences got shorter: long content words fell from 41.7% to 38.6%, the lowest share of any Claude model analyzed, and sentences dropped 17% on average, from 12.14 to 10.03 words.

Semicolons took a hit too. Semicolon usage fell sharply from 6.10 to 1.64 per 1,000 words. The vocabulary is simpler. The phrasing is tighter. If you've been post-processing Claude output to strip AI-sounding punctuation, you can probably retire that script.

How the em dash became AI's scarlet letter

For two solid years, the em dash was the go-to signal for sniffing out AI text. The old em dash trick no longer works: for years, this mark was the smoking gun of text written by ChatGPT or Claude, so models were trained to almost never use it. The irony? The result is paradoxical. Today you can get accused of using AI just because you write clean, well-structured sentences.

Anthropic never marketed this as a feature. The behavioral changes emerged from independent analysis comparing Claude Opus 5.5 output with Claude Opus 5 across matched prompts. No press release. No changelog entry. The model just quietly stopped doing the thing everyone made fun of.

The tradeoff: Claude Opus 5.5 talks more

Fewer tics, sure, but also more words. Answers get 6% wordier, rising from 453 to 481 words on average, and that makes Opus 5.5 the longest-writing model across the Opus family. A new tell may already be forming: hedges and caveats such as "perhaps" and "arguably" rise 97%, from 0.39 to 0.77 per 1,000 words, the highest rate of any Claude model analyzed.

So Claude swapped a visible tic (punctuation) for a subtler one (excessive hedging). The mechanism is an arms race: a tic becomes famous, models learn to avoid it, and whoever hunts artificial text has to find another one. The detective work just got harder.

Beyond style: what Opus 5.5 actually ships

Writing quirks aside, Claude Opus 5.5 launched September 22 with a 20% price cut on Anthropic's flagship model. The new version also makes changes to how Opus communicates, with less jargon and important information placed at the start of messages.

Against OpenAI's GPT-6 Astra and Google's Gemini 3.1 Pro, Anthropic is clearly betting that readability matters as much as raw capability. For teams already tracking Claude's invisible watermarking system, the style cleanup is icing on the cake.

What are the new AI writing tells after the em dash?

The strongest single word signaling "AI wrote this" is now "ensuring," over-represented 4.3× in AI inputs, followed by a family of hedging verbs like "ensures," "highlights," and "reflects." Structural patterns now outweigh punctuation as detection signals.

Did Anthropic officially announce the em dash reduction?

No. The behavioral changes emerged from independent analysis, not from Anthropic. The shift is measurable and real, but it does not appear in any official model documentation or release notes.