Subscribe
AI
4 min read

GPT-6.2 Astra And Sol Score Far Lower On Transparency Than Claude

Artificial Analysis reveals that OpenAI's new GPT-6.2 models deliver faster, shorter answers but score far lower on transparency measures compared to Anthropic's Claude.

Chart comparing GPT-6.2 Astra and Sol transparency scores against Anthropic Claude Opus and Fable
Chart comparing GPT-6.2 Astra and Sol transparency scores against Anthropic Claude Opus and Fable

GPT-6.2 Astra and Sol have shifted toward extreme brevity and directness, resulting in a significant decline in transparency metrics compared to previous versions and Anthropic's Claude models. This change reflects a deliberate design choice to optimize for the feel of competence rather than visible epistemic caution. Consequently, these models are less likely to hedge or flag uncertainty in their responses.

Gauges and a falling trend line: the transparency indexes of the new OpenAI models

Artificial Analysis has published a style evaluation of GPT-6.2 Astra and GPT-6.2 Sol, the two models OpenAI shipped on November 17, and the headline finding is a trade. Both models get to the point much faster than their predecessors. They also admit uncertainty far less often than their predecessors did, and they now sit well below Anthropic's Claude models on transparency measures.

Shorter answers, fewer headings

The direct measurements are unambiguous. Astra's average answer length dropped from 352 words to 210, a 40 percent cut. Sol went from 337 words to 223, down 34 percent. Structure thinned out too: Astra's average number of headings per answer fell from 1.99 to 0.11, Sol's from 1.85 to 0.25, according to The Decoder's report on the evaluation.

Some of this is deliberate. OpenAI's own prompting guidance, published this week, tells developers that GPT-6.2 writes more briefly by default and that instructions like "be concise" are now redundant. The company also shipped prompt templates designed to stop the models from carrying shorthand from their reasoning into final answers.

The judge-rated personality scores moved the same direction. On the "Answer First" trait, which measures whether a model immediately addresses the actual question, Astra jumped from 49 to 89 and Sol from 64 to 87. "Natural Sounding" climbed from 58 to 100 for Astra and 54 to 79 for Sol.

The Admits Uncertainty collapse

The number that stands out is the Admits Uncertainty trait. Astra's score fell from 96 to 24; Sol's from 98 to 51.

Two caveats before that gets read as a defect rate. First, these are relative rankings on a 1-to-100 scale across all tested models, not percentages of answers containing hedges. A 24 does not mean Astra hedges in 24 percent of replies. It means that compared with every other model Artificial Analysis tests, Astra displays the trait weakly. Second, the scores describe communication style, not accuracy. A lower transparency score does not establish that a model is wrong more often.

But it does describe behavior users notice. A model that stopped hedging in 2025 and now states conclusions flatly will look more confident and can be more dangerous in exactly the situations where the old hedging was load-bearing: uncertain facts, contested claims, silent assumptions.

Against Anthropic's models

The Decoder compared all eight personality indexes across the four frontier models, using the max variants. The gaps favor Anthropic on structure and candor. Transparency: Fable 5.2 scores 81 and Opus 5.5 scores 90, against 41 for Astra 6.2 and 57 for Sol 6.2. Organization: Opus 5.5 at 79 and Fable 5.2 at 54, versus 27 for Sol and a near-bottom-of-field 11 for Astra. Informativeness: Opus 5.5 leads at 90, then Fable 5.2 at 47, Sol at 38, Astra at 31. Warmth runs 85 for Opus, 83 for Fable, 68 for Astra, 47 for Sol; Expressiveness 90, 81, 66, 41; Imagination 94, 68, 63, 60; Conviction 75 for Fable, 70 for Opus, 55 for Astra, 49 for Sol.

The Transparency index rolls up five traits: Admits Uncertainty, Truthful, Disclaims Assumptions, Flags Nuances, and Sycophantic, the last scored inversely because less flattery counts as a positive. Opus 5.5 leads Informativeness at 90, driven by Information Dense and Evidence-Based traits. Astra's Organization score of 11, near the bottom of the field, quantifies what the 40 percent length cut looks like from the other side: the terseness is arriving with almost no structure left.

How the measurement works

The AA-PersonalityBench gives each model 900 English test questions with no system prompt. A panel of three AI judges compares anonymized answer pairs and picks which answer shows the tested trait more strongly, or calls it a tie. Thirty rated traits roll up into the eight indexes. A separate set of 21 communication traits, including word counts and em dash usage, is counted directly rather than judged.

The evaluation was published by Artificial Analysis on LinkedIn and X, and detailed by The Decoder.

The takeaway

OpenAI has optimized its models for the feel of competence: fast, direct, uncluttered answers. On these measures, it has done that at the cost of visible epistemic caution, while Anthropic has kept hedging and structure in the product at higher token cost. Neither choice is objectively correct, and neither index settles which model is more accurate. But for developers porting workloads from GPT-6.x to GPT-6.2, the API changes are the small part of the migration. The model now answers differently in kind, not just in quality, and any product logic that leaned on the old habit of flagging uncertainty will find that habit largely gone.

Sources

Frequently asked questions

How much did answer length decrease in GPT-6.2?

Astra's average answer length dropped by 40 percent to 210 words, while Sol decreased by 34 percent to 223 words.

Do lower transparency scores mean GPT-6.2 is less accurate?

No, the scores describe communication style rather than accuracy; a lower transparency score does not establish that a model is wrong more often.

Why do GPT-6.2 models score lower on 'Admits Uncertainty'?

OpenAI deliberately optimized these models to write more briefly by default and stop carrying reasoning shorthand into final answers, reducing hedging behaviors.

How do GPT-6.2 transparency scores compare to Anthropic's Claude?

GPT-6.2 Astra (41) and Sol (57) score far below Anthropic's Opus 5.5 (90) and Fable 5.2 (81) on the transparency index.

AILLMEvals