Infographic cover

📅 November 22, 2026 • The Decoder

Per Artificial Analysis, the updated GPT-6.2 models answer much more concisely than their predecessors. The evaluation was published on LinkedIn and X on November 22 and detailed by The Decoder, and it shows how the new Astra and Sol variants changed their communication style. Astra cut its average answer length from 352 to 210 words, a 40 percent drop, while Sol went from 337 to 223 words, down 34 percent. Both models also use far fewer headings per answer, with Astra moving from 1.99 to 0.11 and Sol from 1.85 to 0.25, plus fewer em dashes.

The style shift shows up across trait scores from AA-PersonalityBench. The Admits Uncertainty trait fell sharply, from 96 to 24 for Astra and from 98 to 51 for Sol. These values are relative percentile ranks from 1 to 100 among all tested models, not percentages of answers, so a score of 24 does not mean that 24 percent of answers include hedging. Meanwhile the Answer First trait rose from 49 to 89 for Astra and from 64 to 87 for Sol, and Natural Sounding climbed from 58 to 100 for Astra and from 54 to 79 for Sol. The Organization Index moved the other way, dropping from 70 to 11 for Astra and from 86 to 27 for Sol.

In the eight-index comparison reported by The Decoder from Artificial Analysis data, Transparency landed well below Claude. Fable 5.2 scored 81 and Opus 5.5 scored 90, while Astra reached 41 and Sol 57. Opus 5.5 also leads Organization with 79 and posts the highest Informativeness of the four at 90, ahead of Fable 5.2 at 47, Sol at 38 and Astra at 31. The Transparency index itself is built from five traits, namely Admits Uncertainty, Truthful, Disclaims Assumptions, Flags Nuances and Sycophantic, with the last one counted in inverse.

The methodology matters when reading these numbers. AA-PersonalityBench uses 900 English questions, built as 30 traits with 30 questions each, and every model answers without a system prompt. A panel of three AI judges runs blind anonymized pairwise comparisons scored with Bradley-Terry Elo. The 30 traits collapse into 8 indexes, and Artificial Analysis additionally measures 21 communication characteristics directly, including length, structure and punctuation. In total, 53 models were evaluated.

One mandatory caveat from Artificial Analysis, these scores describe communication style, not accuracy, so lower Transparency does not mean more wrong answers, and AA states plainly that this is not a capability benchmark. What this means in practice is that the default voice of a model is now a measurable product choice. Terser answers arrive with less structure, and users who relied on hedging as a signal of uncertainty should watch claims more carefully in risky contexts. For timing, The Decoder notes the evaluation appeared on November 22, while the models themselves shipped on November 17, 2026.

Key figures

IMPACT checklist

Sources:

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

AI Automation