Claude Models Lead Separate Writing Traits in AI Benchmark

Claude Fable 5.1 and Claude Haiku 5.5 topped different writing-style measures in a 40-model AI benchmark, while a separate combined score put Muse Spark 1.3 first.

At 8:34 p.m. ET Oct. 22 (00:34 UTC Oct. 23), the benchmark listed Fable first for Information Dense and Haiku first for Literate. Information Dense tracks how much specific detail responses carry; Literate reflects grammar, punctuation and word choice. The snapshot covered 40 models, so the rankings reflect that leaderboard at one moment.

The labels describe different dimensions. Muse Spark 1.3 led Informativeness, a combined index drawing on Information Dense, Evidence-Based, Resource-Specific, Actionable Guidance and Pragmatic measures. Literate belongs to Expressiveness, while Information Dense also contributes to Informativeness. A trait ranking and a combined-index result answer different questions; they do not establish an overall winner.

AA-PersonalityBench debuted Oct. 19. It covers 30 writing traits, with 30 questions for each trait, for 900 questions total. The questions are grouped by trait rather than combined into one undifferentiated writing exam. Eight indexes organize those traits, and 21 communication characteristics, such as response length and structure, are measured separately. Each model answers every question once, without a system prompt.

Each prompt is presented on its own as a user message. In each comparison, one judge evaluates two anonymized answers to the same question, with the order randomized. The judge is selected from Gemini 3.8 Flash, Claude Sonnet 5.5 and GLM 5.3 Flash. A Firth-penalized Bradley-Terry model aggregates the head-to-head judgments into relative scores.

The scores measure how prominently a writing trait appears, without grading factual accuracy or whether that trait is desirable. They also do not show that one model is more capable overall.

The writing results add style measures to earlier coverage of the benchmark’s sycophancy rankings.