Claude Opus 5.5 Moves Closer to Human Writing but Retains 2,548 Tells

Graphite compared articles from ten models and human writers across 9,974 matched topics. [1] Claude Opus 5.5 recorded a word-distribution divergence score of 0.052 against the human sample, down from 0.064 for Opus 5, a 19% reduction. [1] Graphite found 2,548 words, phrases and sentence patterns that appeared at least twice as often in Opus 5.5 articles as in comparable human articles. [1]

The model produced 0.015 em-dashes per 1,000 words, versus 2.92 for Opus 5, a decline of about 99%. [1] OpenAI’s GPT-6 Astra scored 0.109 on word-distribution divergence, up from 0.101 for GPT-5.6 Sol. [1]

Claude Opus 5.5’s lower divergence score indicates closer alignment with Graphite’s human sample, but the remaining 2,548 tells show why a single stylistic cue is an inadequate test. The near-disappearance of em-dashes removes one conspicuous habit without making the model’s broader language patterns indistinguishable from human writing.

The comparison with GPT-6 Astra also shows that newer generations do not necessarily move in the same direction on this measure: Opus moved closer to the human sample while Astra’s divergence increased. What the sources don't address: whether these measured differences remain consistent across languages, formats and real enterprise writing workflows.