Claude's writing has been changing under the hood, and now there's data to prove it. Arena.ai analyzed tens of thousands of real-world, high-reasoning outputs from Claude Opus 4.5 and Opus 5 collected through its Text Arena. Opus 5 writes more, structures sentences more elaborately, and uses simpler vocabulary than its predecessor.

What Arena actually is

Arena (formerly Chatbot Arena) is a public, web-based platform that evaluates large language models. Users enter prompts for two anonymous models to respond to and vote on the better response, after which the models' identities are revealed. As of early 2026, Arena has over 5 million monthly active users across 150 countries and processes more than 60 million conversations per month. That scale means the writing data being pulled from is genuinely representative of how these models behave in the wild, not in a lab.

Longer, more elaborate, but simpler words

The headline stat is hard to ignore. Opus 5 averages 510 words per response, compared to 158 for Opus 4.5, a 3x increase in output length. The shift goes deeper than word count:

• 58% rise in average sentence length from Opus 4.5 to Opus 5
• 46% rise in clause frequency (how often sentences contain embedded sub-clauses)
• 6.8 percentage point drop in long content words (from 46.9% to 40.1%)
• 53% drop in abstract nouns (from 6.02 to 3.79 per 1,000 words)

That last pair is the surprising twist. Opus 5 builds longer, more structurally complex sentences, then fills them with simpler, more concrete words. The model has become more elaborate in structure while becoming more accessible in vocabulary. Whether that reflects a deliberate design choice or an emergent artifact of training is an open question.