Anthropic's newest model nearly abandoned the em dash, but Graphite counted 2,548 statistical tells that still give it away
Anthropic’s newest model has learned to write more like a person. It just can’t stop calling things dependable.
A study from Graphite, shared on September 30, 2026, found that Claude Opus 5.5 uses the word “dependable” 23 times more often than human writers do. That is one of 2,548 statistical tells Graphite identified in the model’s output. Each one is a quirk that machines produce far more often than people.
The em dash has left the building #
The study arrived barely a week after launch. Anthropic released Opus 5.5 on September 22, 2026, and pitched it around clearer, more natural communication with less verbosity.
On at least one measure, that pitch holds up. Graphite reported that em dash usage fell approximately 99% compared with Opus 5, dropping from 2.92 to 0.015 per 1,000 words.
The broader numbers also moved in the right direction. Graphite measured word-distribution divergence, which compares how often a model reaches for particular words against how often humans do. A lower score means the model’s vocabulary habits look closer to ours.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Opus 5.5 scored 0.052, down from 0.064 for Opus 5. Graphite describes that as a 19% improvement. Put simply, the new model’s word choices drift less from typical human patterns than its predecessor’s did.
The study also recorded a 37% decrease in mannered prose.
Better disguise, same fingerprints #
Despite all of those improvements, Graphite found that overall detection strength remains largely unchanged.
The 2,548 tells in Graphite’s count are features that show up at least 2x more often in Opus 5.5 text than in human writing. “Dependable” is the headline example, at 23 times the human rate.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our