Claude Opus 5.5 appears to write with fewer traces of AI than the familiar patterns often attributed to artificial intelligence models. New Arena analysis shows that it uses far fewer hyphens, shorter sentences, and simpler vocabulary than Opus 5.
The change is of interest to those who leverage AI tools in articles, reports, and professional communications. It doesn't mean that a text automatically becomes human or free of stereotypes, but it does show that companies are now improving the way models express themselves.
See also: Autonomous AI Hacking: Who is Legally Responsible?
Fewer AI traces in Claude Opus 5.5
According to data presented by Arena and reported by BleepingComputer, the comparison was based on high-stakes responses from Text Arena, which were collected in August and September 2026. Out of 12 writing metrics, 10 moved in the direction that Arena considers improved.
The most striking difference concerns the em dash. Its frequency fell from 15.2 to 0.8 occurrences per 1,000 words, or about 95%. At the same time, colons decreased from 6.10 to 1.64 per 1,000 words, while the average sentence became shorter, from 12.14 to 10.03 words.
There is, however, an interesting contradiction. Although the sentences became shorter, the average response grew from 453 to 481 words. The model thus appears less “mechanical” in its syntax, but not necessarily shorter. The improvement concerns mainly the pace and word choice.
Why natural writing matters
Dashes, overly long sentences, and repetitive paragraph organization are not proof that a text was written by AI. However, they can make a text predictable or create the impression of automated production, especially when combined with general expressions.
Anthropic claims in its presentation that Claude Opus 5.5 puts the most important point earlier, uses less jargon, and follows more closely the style rules given by the user. These are statements by the company and not independent evidence for any type of work.
Claude Opus 5.5 was introduced on September 22 as the first model in the new Claude 5.5 family. The company positions it for long-running programming and knowledge tasks, while the API price is $4 per million input tokens and $20 per million output tokens. For Opus 5, the corresponding prices were $5 and $25.
Lower price and more consistent guidance can help teams using AI for drafting, technical documentation, or customer service. But they don't solve the problem of reliability. A model that sounds more natural can also produce more convincingly wrong information.
See also: Meta Muse: Communication with the AI agent via video calls and email
What users should keep in mind
The change does not eliminate the need for human editing. Arena did not disclose the sample size, exact questions, model settings, or statistical significance of the differences in this analysis. Therefore, the percentages describe the behavior of the specific set of responses and are not a universal rule.
Pangram's independent evaluation , which examined 1,000 common queries in Opus 5.5, Opus 5, and Sonnet 5, also concludes that the two Opus models write differently. The same study notes that some indicator phrases have decreased, while others, such as example introductions and summary transitions, have appeared more frequently. The findings are specific to the specific sample and are not a detection tool in themselves.
For professionals, the practical benefit is more consistent adherence to a style guide and less need for mechanical corrections. Fewer AI footprints make the result smoother, not automatically more accurate. For readers, every important claim needs verification, source, and human judgment.
See also: Anthropic and OpenEvidence offer free medical AI
Claude Opus 5.5 shows that the evolution of models is not only measured in performance and cost. It is also measured in their ability to communicate clearly, without repetitive patterns. The final quality, however, still depends on the request, the control of the answers and the person who signs the result.
