AI Tells: Opus 5.5 Update

Introduction

On September 16th, 2026, we published our research on AI tells. On September 22nd, Anthropic released Claude Opus 5.5. In this paper, we extend our analysis to include Opus 5.5.

The tells data and raw article data are available to download.

Opus 5.5 Still Has Many Tells

We identify words, phrases, and frames (word patterns with underscores representing gaps of up to three words) that occur at least twice as often in AI writing as in human writing, after normalizing for the amount of text and applying frequency thresholds. We also compare stylistic features, such as em-dash use and variation in sentence length. We use the same methods as in our original study.

We find 2,548 word, phrase, and frame tells in Opus 5.5, just 4% fewer than the 2,666 we find in Opus 5.

The examples below show recurring themes in Opus 5.5’s tells.

Tells by Model

The explorer below shows the most disproportionately used words, phrases, frames, and features, along with LLM-summarized themes. For the stylistic features, the larger the absolute Cohen's d, the more strongly the feature distinguishes AI from human writing.

Opus 5.5 Has Different Tells Than Astra

We next compare Opus 5.5 directly with Astra.

More Qualification in Astra

Compared with Claude Opus 5.5, GPT-6 Astra qualifies claims more often.

More Superlatives in Opus 5.5

Compared with GPT-6 Astra, Claude Opus 5.5 uses more superlatives.

More Differences Between Models

Below, we show the most disproportionately used words, phrases, frames, and features for Opus 5.5 and Astra, along with LLM-summarized themes.

Opus 5.5 Has Different Tells Than Opus 5

More Language About Helpfulness

Opus 5.5 uses more language about helping and making things easier.

Less Emphatic Phrasing

Opus 5.5 uses emphatic phrases less often than Opus 5.

Less Mannered Prose

Mannered prose substitutes metaphor and flourish for direct statements. We use Claude Opus 5 to score mannered prose on a scale from 0 to 100 in a subsample of 1,000 matched topics. Opus 5.5’s mannered-prose score is 37% lower than Opus 5’s, at 10.57 compared with 16.75. It remains about 1.6 times the human score, compared with 1.2 times for Astra.

Em Dashes Nearly Disappear

Opus 5.5 uses 0.015 em dashes per 1,000 words, compared with 2.92 for Opus 5, a decrease of 99%. Both Opus 5.5 and Astra use em dashes much less often than human writers. Astra uses them about one-eighth as often, while Opus 5.5 has nearly stopped using them.

More Changes Between Model Versions

Below, we show the words, phrases, frames, and features that differ most between Opus 5 and Opus 5.5, along with LLM-summarized themes.

Opus 5.5’s Writing Is Becoming More Similar to Human Writing

These measures capture different aspects of writing. Mannered prose and well-known tells describe specific writing habits. Divergence measures the overall word distribution, while the tell count measures how many words, phrases, and frames are disproportionately common. A model can therefore reduce familiar tells without bringing its overall word use closer to human writing, or move closer overall while retaining thousands of tells.

Well-Known Tells Are Becoming Less Prominent

We track the same 11 features as in our original study, including words like “delve,” formulaic closes, and hedging. We measure how much these features differ between AI and human writing, accounting for variation across articles.

The average strength of these 11 tells is 6% lower in Opus 5.5 than in Opus 5 and 53% lower than in Opus 4. For GPT, it is 46% lower in Astra than in GPT-4.1, but changes little after GPT-5.

Opus 5.5 Still Has Thousands of Tells

The number of tells falls from 3,746 in Opus 4 to 3,059 in Opus 4.6, 2,666 in Opus 5, and 2,548 in Opus 5.5. Despite this decline, thousands of words, phrases, and frames remain disproportionately common in Opus 5.5’s writing.

Claude’s Word Distribution Is Becoming More Similar to Human Writing

We use Jensen-Shannon divergence to compare probability distributions over words and phrases. Larger values mean greater differences in how the model and human writers use words and phrases.

Opus 5.5’s word distribution is closer to human writing than Opus 5’s. Its unigram divergence falls from 0.064 to 0.052. For GPT, divergence rises from 0.101 for Sol to 0.109 for Astra. Opus 5.5 is also closer to human writing than Opus 4, while Astra is further away than GPT-4.1.

These trends are not driven by a few common words. Claude’s divergence falls from Opus 4 to Opus 5.5, and GPT’s rises from GPT-4.1 to Astra, even after removing the 100 most common words in human writing. The difference also extends beyond common function words. Nouns, verbs, adjectives, and adverbs together account for about 79% of the divergence gap between Astra and Opus 5.5.

Methodology

We use the same article-generation and analysis methods as in our original AI Tells study, adding Claude Opus 5.5.

We analyze n-grams and frames across 9,974 topics with both a human article and an article from every model. Adding Opus 5.5 changes this shared set slightly, so results for earlier models may differ from the original study.

Limitations

The limitations of our original study also apply to this update.

Conclusion

Tells shift between versions, but they do not go away. Changes intended to improve a model’s writing may reduce some tells while introducing or amplifying others. Opus 5.5 uses more language about helping and making things easier than Opus 5, while emphatic phrasing and em dashes become less common. Its mannered-prose score is also 37% lower than Opus 5’s, though still higher than Astra’s.

Opus 5.5 also has different tells from Astra. It uses more superlatives, while Astra qualifies claims more often.

Well-known tells are less prominent in Opus 5.5, and its overall word distribution is closer to human writing than Opus 5’s. In contrast, Astra’s word distribution is further from human writing than Sol’s.