Anthropic’s Claude Opus 5.5 is quietly rewriting the stylistic DNA of AI-generated text, effectively killing one of the most recognizable markers of machine writing: the em dash. According to new analysis from Arena, an AI benchmarking tool, the model’s usage of em dashes has plummeted by approximately 95% compared to its predecessor, Opus 5. This drastic reduction signals a deliberate shift in Anthropic’s approach to natural language generation, moving away from the 'AI slop' aesthetic that has plagued LLM outputs for years.

The Death of the Em Dash

The data is stark. Opus 5 averaged 15.2 em dashes per 1,000 words, a frequency that immediately flagged text as AI-generated to many human readers. Opus 5.5 slashed this to just 0.8 per 1,000 words. Semicolon usage followed a similar downward trend, dropping from 6.10 to 1.64 per 1,000 words. These punctuation changes are not random; they represent a concerted effort to reduce the syntactic patterns that betray an LLM’s hand. Arena analyzed high-reasoning Text Arena responses from August and September 2026, finding that 10 of 12 writing measures moved in a direction considered 'better' for human-like readability.

Shorter Sentences, Longer Answers

While the punctuation is cleaner, the structure has changed in ways that may frustrate users seeking conciseness. Opus 5.5 writes significantly shorter sentences, averaging 10.03 words per sentence compared to 12.14 for Opus 5. This simpler wording contributes to a more natural flow, but it comes at a cost. The model has become more verbose overall, with average answer lengths increasing from 453 to 481 words. This makes Opus 5.5 the longest-writing version in the current Opus comparison, suggesting that Anthropic is prioritizing thoroughness and natural pacing over brevity.

Key Takeaways

  • Claude Opus 5.5 reduced em dash usage by ~95% compared to Opus 5 (0.8 vs 15.2 per 1,000 words).
  • Semicolon usage dropped from 6.10 to 1.64 per 1,000 words, further reducing AI-typical punctuation.
  • Average sentence length decreased from 12.14 to 10.03 words, promoting simpler syntax.
  • Overall response verbosity increased, with average word counts rising from 453 to 481.
  • Arena’s analysis covers high-reasoning responses from August and September 2026.

The Bottom Line

Anthropic is finally tuning Claude to sound less like a robot and more like a writer, even if it means we all have to read a few more words to get the point.