OpenAI's New Astra Model Writes 40% Shorter, but Its Hedging Score Fell from 96 to 24
OpenAI made GPT-6.2 Astra and Sol shorter and more direct, but the models less often say "I'm not sure." Artificial Analysis found that the average Astra response shrank from 352 to 210 words (−40%), and Sol's from 337 to 223 (−34%). Astra's admits-uncertainty rank (its place among models on a 1–100 scale) fell from 96 to 24, and Sol's from 98 to 51. That's why, for medicine, law, and finance, you should explicitly require caveats in your prompt.
AI-processed from The Decoder; edited by Hamidun News
Many people have complained that chatbots say too much: three paragraphs of preamble, lists, headings, and reflexive disclaimers. OpenAI seems to have listened. Analysts at Artificial Analysis compared the new GPT-6.
2 Astra and Sol with their predecessors and found two things. First, responses got 34–40% shorter: Astra now averages 210 words instead of 352, and Sol 223 instead of 337. Second, the models much less often say "I'm not sure."
The ranking for this trait (the model's place among all tested models, on a scale from 1 to 100) dropped for Astra from 96 to 24, and for Sol from 98 to 51. Picture an assistant that used to say "this seems right, but double-check it" and now just answers in a confident tone. For those building products on these models, this matters more than it might seem.
What changed in response style
Artificial Analysis ran the models through its own AA-PersonalityBench test and separately measured simple metrics like response length. Astra's average response fell from 352 to 210 words, a 40% drop. Sol went from 337 words to 223, down 34%.
The layout changed too. Astra averaged 1.99 headings per response, now 0.11. Sol had 1.85, now 0.25. Em dashes, which many have learned to recognize as a sign of AI-generated text, also became less frequent.
The organization index (structure, clarity, logical flow) dropped: Astra's fell from 70 to 11, and Sol's from 86 to 27. But the Answer First trait, which measures how directly a model answers the question asked, rose: Astra's went from 49 to 89, and Sol's from 64 to 87. The naturalness score of the speech also increased: from 58 to 100 for Astra and from 54 to 79 for Sol.
The cost of brevity: fewer honest caveats
The sharpest drop occurred in the Admits Uncertainty trait, meaning the willingness to acknowledge that a model doesn't know something or isn't certain.
An important caveat: these are not percentages of responses. The scores show a model's relative place among all tested models on a 1 to 100 scale. A score of 24 does not mean Astra admits uncertainty in 24% of cases. It means that on this trait it ranks in the lower part of the leaderboard.
The composite transparency index includes five traits: admitting uncertainty, truthfulness, caveats about assumptions, pointing out nuances, and absence of flattery (for the last one, less flattery is better). Astra scores 41 and Sol 57. Anthropic scores higher: Claude Fable 5.2 earned 81 and Opus 5.5 earned 90. Anthropic also leads on organization: Opus 5.5 got 79, Fable 5.2 got 54. On informativeness, Opus 5.5 leads with 90 points against 31 for Astra, 38 for Sol, and 47 for Fable 5.2.
How to read this ranking and what developers should do
The test works like this: 900 English-language questions, with models answering without a system prompt. A panel of three AI judges looks at anonymous pairs of responses and decides which one expresses the given trait more strongly, or calls it a tie. From this come 30 traits and eight composite indexes. The comparison was run for variants with maximum reasoning levels.
The main limitation: this measures style, not quality. A low transparency index does not prove that a model makes more mistakes. It only shows that it less often accompanies answers with caveats. Moreover, the judges are themselves neural networks, and the questions were only in English, so the picture may differ for Russian-language tasks.
The practical takeaways are these:
- For chat interfaces and help bots, where users want a quick answer, OpenAI's new style can be a plus: the text is shorter, uses fewer tokens, and is faster to read.
- For medicine, finance, legal, and analytical scenarios, where the degree of confidence matters, explicitly ask the model to state its assumptions and level of certainty. Without such a request, the test suggests it will sound more confident than it actually is.
- If you are migrating from the previous version, recheck your prompts and parsers. The disappearance of headings and lists can break formatting that used to work on its own.
- When choosing between OpenAI and Anthropic, look beyond benchmark accuracy to presentation style. It affects how users trust answers and verify them.
Bottom line: OpenAI has bet on a concise, conversational style, while Anthropic remains, on these measures, closer to cautious and elaborate. Which approach is better depends on the product. It is worth testing the difference on your own data before rolling out to production.
- Nov 22, 2026Five Free GitHub Repos Open 1,400+ AI Models Behind One Endpoint
- Nov 22, 2026OpenAI Cut the Success Rate of Reasoning-Theft Attacks from 54% to 0.2%
- Nov 22, 2026Microsoft's Agent Hand-Wrote Neural Network Weights and Hit 100% Accuracy on 13 Tasks
- Nov 22, 2026LangChain, a Side Project Turned $3 Billion Startup in Four Years, Raises $300 Million
Key Questions Answered
What changed in response style?
Artificial Analysis ran the models through its own AA-PersonalityBench test and separately measured simple metrics like response length. Astra's average response fell from 352 to 210 words, a 40% drop. Sol went from 337 words to 223, down 34%. The layout changed too. Astra averaged 1.99 headings per response, now 0.11. Sol had 1.85, now 0.25. Em dashes, which many have learned to recognize as a sign of AI-generated text, also became less frequent.
What are the key details on The cost of brevity: fewer honest caveats?
The sharpest drop occurred in the Admits Uncertainty trait, meaning the willingness to acknowledge that a model doesn't know something or isn't certain. An important caveat: these are not percentages of responses. The scores show a model's relative place among all tested models on a 1 to 100 scale. A score of 24 does not mean Astra admits uncertainty in 24% of cases. It means that on this trait it ranks in the lower part of the leaderboard.
How to read this ranking and what developers should do?
The test works like this: 900 English-language questions, with models answering without a system prompt. A panel of three AI judges looks at anonymous pairs of responses and decides which one expresses the given trait more strongly, or calls it a tie. From this come 30 traits and eight composite indexes. The comparison was run for variants with maximum reasoning levels. The main limitation: this measures style, not quality. A low transparency index does not prove that a model makes more mistakes.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.