OpenAI shortens its responses, at the cost of less transparency
An evaluation by Artificial Analysis compares GPT-6.2 Astra and Sol to Claude Fable 5.2 and Opus 5.5. OpenAI’s models respond more briefly but admit uncertainty less often.
According to an evaluation by Artificial Analysis published on LinkedIn and X, OpenAI’s GPT-6.2 Astra and Sol models respond faster and more directly than their predecessors but lose ground on transparency compared to Anthropic’s Claude models. Astra and Sol produce answers 34 to 40% shorter, with far fewer headings and em dashes. In contrast, Claude Fable 5.2 and Opus 5.5 maintain the advantage in organization and dense information. The debate is not cosmetic: it reveals something about how AI assistants choose between conciseness and caution.
Des réponses plus courtes, donc plus tranchantes
GPT-6.2 Astra drops from 352 to 210 words on average, a 40% decrease, while Sol falls from 337 to 223 words, a 34% drop. At the same time, Astra goes from 1.99 to 0.11 headings per response and Sol from 1.85 to 0.25. Both models therefore take fewer detours, fewer subsections, and give the impression of aiming for the useful answer rather than a small academic exposition.
The change is also seen in the “Answer First” index, which measures whether the model answers the question immediately. Astra jumps from 49 to 89 and Sol from 64 to 87. Their language also appears more natural according to Artificial Analysis, with scores rising from 58 to 100 for Astra and from 54 to 79 for Sol. In short, OpenAI clearly asked its models to speak more directly. They understood the brief, without rambling. An almost suspicious rarity in the chatbot world.
Mais la prudence s’efface aussi
The clearest trade-off concerns the ability to recognize uncertainty. On the “Admits Uncertainty” trait, Astra falls from 96 to 24 and Sol from 98 to 51. However, note: these scores are relative, on a scale from 1 to 100, and do not mean that a model “admits uncertainty” in 24% of its answers. They rather indicate its position relative to other tested models. The nuance matters, especially when dealing with systems meant to assist decision-making, not just to produce neatly arranged text.
Artificial Analysis specifies that these scores measure communication style, not overall answer quality. A drop in the transparency index therefore does not mean a model is more often wrong. However, it suggests it exposes its reservations less, nuances its statements less, and allows itself to get straight to the point more. For the user, this can give an impression of clarity. For professional use, it can also hide blind spots.
Claude garde l’avantage sur la structure et la densité
The comparison with Claude Fable 5.2 and Opus 5.5 is harsh for OpenAI. In the Artificial Analysis table, Fable scores 81 in transparency and 54 in organization, versus 41 and 11 for Astra, and 57 and 27 for Sol. Opus 5.5 does even better with 90 in transparency, 79 in organization, and 90 in informativeness. It is the highest overall score on this index, notably thanks to traits like “Information Dense” and “Evidence-Based.”
In other words, Anthropic maintains a clear lead here in the ability to structure rich answers without turning them into a wall of text. The two approaches diverge sharply: OpenAI pushes for more concise and direct outputs, while Anthropic scores better on caution, hierarchy, and information density. For companies integrating these models into customer support, knowledge bases, or internal tools, the choice is therefore not just about raw performance. It concerns tone, level of detail, and how the AI signals its limits.
Key points
- Astra reduces its responses by 40% according to Artificial Analysis.
- Sol decreases by 34% in average length.
- Opus 5.5 scores 90 in informativeness.
- The measurement is based on 900 questions in English.
- Anthropic leads in transparency and organization.
- The scores compare “max” variants of each model.
By the numbers
- 900 questions — basis of the AA-PersonalityBench benchmark, in English, according to Artificial Analysis.
- 40% — decrease in Astra’s average response length.
- 34% — decrease in Sol’s average response length.
- 90 points — Opus 5.5’s score on the informativeness index.
- 1 to 100 — relative scale used for personality indices.