OpenAI’s latest model variants, GPT-6.2 Astra and Sol, have fundamentally shifted their communication style toward extreme conciseness, according to new data from Artificial Analysis. Published on November 22, the evaluation highlights a drastic reduction in answer length and structural complexity compared to predecessors, but this gain in brevity comes at a significant cost to transparency. While the models now deliver answers 34-40% shorter, they score substantially lower on transparency metrics than Anthropic’s Claude models, suggesting a deliberate product choice that prioritizes speed and directness over explanatory depth.
The Shift to Concise Answers
The data shows a clear departure from the verbose outputs typical of earlier LLMs. Astra’s average answer length dropped from 352 to 210 words, representing a 40% decrease, while Sol’s length fell from 337 to 223 words, a 34% reduction. This conciseness is accompanied by a near-total elimination of structural elements; Astra’s use of headings per answer plummeted from 1.99 to 0.11, and Sol’s dropped from 1.85 to 0.25. The models also reduced their use of em dashes, signaling a move toward a more streamlined, less segmented presentation of information.
Transparency Scores Lag Behind Claude
In the eight-index comparison reported by The Decoder, the transparency gap between GPT-6.2 variants and Anthropic’s offerings is stark. Claude Opus 5.5 scored 90 on the transparency index, and Fable 5.2 scored 81. In contrast, GPT-6.2 Sol reached only 57, while Astra scored a mere 41. The transparency index aggregates traits like Admits Uncertainty, Truthful, Disclaims Assumptions, Flags Nuances, and Sycophantic (inverted). The sharp decline in the 'Admits Uncertainty' trait is particularly notable, with Astra’s score falling from 96 to 24 and Sol’s from 98 to 51. It is crucial to note that these are relative percentile ranks among 53 tested models, not percentages of answers containing hedging.
Methodology and Context
The evaluation relies on AA-PersonalityBench, which uses 900 English questions across 30 traits, with each model answering without a system prompt. A panel of three AI judges conducts blind anonymized pairwise comparisons using Bradley-Terry Elo scoring. While Opus 5.5 leads in organization and informativeness, the GPT-6.2 variants show improvements in 'Answer First' and 'Natural Sounding' traits. However, Artificial Analysis explicitly states these scores describe communication style, not accuracy. A lower transparency score does not imply the model is wrong, but it does mean the model is less likely to explain its reasoning or acknowledge uncertainty.
Key Takeaways
- GPT-6.2 Astra and Sol reduced average answer length by 40% and 34% respectively compared to predecessors.
- Transparency scores for Astra (41) and Sol (57) significantly trail Claude Opus 5.5 (90) and Fable 5.2 (81).
- The 'Admits Uncertainty' trait dropped sharply for both variants, with Astra falling from 96 to 24.
- Structural elements like headings were nearly eliminated, with Astra’s usage dropping from 1.99 to 0.11 per answer.
- These metrics reflect communication style choices, not factual accuracy or capability benchmarks.
The Bottom Line
OpenAI is clearly optimizing for a product voice that feels more human and less robotic, but users relying on LLMs for critical decision-making need to be wary. When a model stops explaining its assumptions and hedging its bets, you are no longer getting a transparent partner, but a concise oracle that demands higher trust and verification from the user.