Anthropic's Claude Opus 5.5 shows improved human-like prose in study

A study by the agency Graphite indicates that Anthropic’s new AI model, Claude Opus 5.5, produces text that more closely mimics human prose compared to its predecessor, Opus 5. The research utilized 9,974 subjects, comparing human-written articles from the pre-ChatGPT era with those generated by various AI models.

Key findings include a 19% reduction in the Jensen-Shannon divergence for Opus 5.5, suggesting a word distribution more faithful to human authors. Additionally, the model showed a 37% decrease in “mannered prose”—defined as the use of metaphors and flourishes instead of direct statements. The use of em dashes also dropped by 99% in the new model, reaching levels lower than those found in pre-ChatGPT human articles.

In contrast, OpenAI’s models showed a widening gap in similarity to human writing, with GPT-6 Astra showing a higher divergence than GPT-5.6 Sol. Despite these improvements in naturalness, researchers noted that thousands of subtle linguistic markers still distinguish AI-generated text from human writing.