Anthropic released the Claude Opus 5.5 model on September 22, 2026, aiming for more natural and concise communication. A study published by Graphite on September 30, 2026, analyzed the model's output to determine how closely its writing patterns resemble human text.
Graphite identified 2,548 statistical markers, defined as features appearing at least twice as often in the model's text as in human writing. The study highlights that the model uses the word dependable 23 times more frequently than human authors.
The research indicates a significant reduction in the use of em dashes, which dropped by approximately 99% compared to the previous version, falling from 2.92 to 0.015 per 1,000 words.
The model showed a 19% improvement in word-distribution divergence, with its score decreasing from 0.064 in Opus 5 to 0.052 in Opus 5.5. Despite these changes, Graphite concluded that the overall ability to detect the model's text remains largely unchanged.
