GPT-6.2 Astra and Sol answer faster and shorter but score lower on transparency than Claude
An Artificial Analysis evaluation finds that OpenAI’s GPT-6.2 Astra and Sol now answer more briefly and more directly than their predecessors, while ranking far lower on willingness to admit uncertainty. Astra’s average length fell from 352 to 210 words (40 percent); Sol’s from 337 to 223 (34 percent). Headings dropped sharply as well—Astra from 1.99 per answer to 0.11, Sol from 1.85 to 0.25—and both models use far fewer em dashes. The Organization Index, which combines structure, clarity and related traits as scored by AI judges, fell from 70 to 11 for Astra and from 86 to 27 for Sol. At the same time Answer First jumped (Astra 49 to 89, Sol 64 to 87), and Natural Sounding language rose (Astra 58 to 100, Sol 54 to 79).
The steepest change is Admits Uncertainty, a relative 1–100 ranking among tested models, not a share of answers: Astra 96 to 24, Sol 98 to 51. Eight personality indexes roll up 30 traits from a panel of three AI judges comparing anonymized pairs; comparisons use “max” variants. Anthropic’s Claude Fable 5.2 and Opus 5.5 score 81 and 90 on the Transparency Index (Admits Uncertainty, Truthful, Disclaims Assumptions, Flags Nuances, and inversely scored Sycophantic) versus Astra 41 and Sol 57, and they also lead on Organization. Opus 5.5 leads Informativeness at 90, with Information Dense and Evidence-Based among its strongest traits.
AA-PersonalityBench uses 900 English questions with no system prompt. Twenty-one communication traits, including length and headings, are counted directly rather than judged. The scores describe style, not overall quality: a lower Transparency Index does not mean more wrong answers. Artificial Analysis posted the evaluation on LinkedIn and X.
Continue in the app — vote & join in ➔