Artificial Analysis's AA-PersonalityBench shows changes in GPT-6.2 Sol's response length, formatting, tendency to answer first, natural-sounding quality, and willingness to admit uncertainty compared with its predecessor. These are communication-style metrics, not measures of factual accuracy or overall capability.
Tasks suited to this evaluation: Assessing communication style, including response length, structure, directness, naturalness, and expression of uncertainty.
Tasks these results should not be generalized to: Factual accuracy, coding ability, overall quality, or prompts, languages, and reasoning-effort levels that were not tested.
Applicable model version: GPT-6.2 Sol. The AA page lists the model variant as GPT-6.2 Sol (Max). The comparison baseline is described only as a “previous version” in the original post, with no specific version number.
Test environment or client: Artificial Analysis's AA-PersonalityBench. The specific API, reasoning parameters, and run date are not provided.
Reasoning effort and parameters: Some comparisons are explicitly from low to max effort. The effort level for the other style metrics is not specified. Generation parameters such as temperature are not provided.
Artificial Analysis distinguishes style tests from capability tests. Its methodology page says each model answers 900 questions once, with each question sent as a single user message and no system prompt. The question set consists of 30 questions for each of 30 personality traits. The full set is not public. Each trait is evaluated through blind pairwise comparisons by model judges, with each comparison assigned to one judge. The judging panel consists of Gemini 3.8 Flash (high), Claude Sonnet 5.5 (medium), and GLM 5.3 Flash (max).
The judgments are used with a Bradley–Terry model to estimate an Elo score for each trait, which is then converted to a cross-model percentile from 1 to 100. Another 21 communication characteristics are calculated directly from the 900 responses. The AA methodology page also says GPT-6.1 Sol (max) is used as the trait Elo anchor (1000). However, the original post does not specify the exact version or effort level for all “previous version” comparisons.
Methodology: AA-PersonalityBench; methodology page. The original post links to both pages, which are included here only to explain the evaluation methodology. The X post listed in the Source section and its updates in the same thread are the sole primary sources for this article.
The following figures are reported for Sol in the original post. Except for words per response and the number of headings and em dashes, the trait figures should be read as cross-model percentiles according to the AA methodology page. The specific predecessor version and some effort levels are not identified in the original post.
| Metric | Change for GPT-6.2 Sol | Measure |
|---|---|---|
| Words per response | 337 → 223 (about 34% fewer) | Average words per response |
| Headings per response | 1.85 → 0.25 | Average count |
| Em dashes per response | 3.01 → 0.51 | Average count |
| Organization Index | 86 → 27 | Percentile |
| Answer First | 64 → 87 | Percentile |
| Admits Uncertainty | 98 → 51 | Percentile |
| Natural Sounding | 54 → 79 | Percentile |
| Evidence-Based, low → max effort | 16 → 68 | Percentile |
| Disclaims Assumptions, low → max effort | 26 → 77 | Percentile |
The original post and updates in the same thread explicitly report the figures in the table. “Less organized but more likely to answer first” refers to two different metrics: Organization Index decreases, while Answer First increases.
The 900 questions, single-user-message format, absence of a system prompt, judging method, judge models, and scoring and percentile methodology come from the AA methodology page. It says the full question set is private.
The AA page distinguishes response-style metrics from correctness. These scores do not indicate that answers are more or less correct.
These results support a limited conclusion: on AA's question set and under its testing process, GPT-6.2 Sol gives shorter answers with less explicit formatting, scores higher on answer-first tendency and natural-sounding quality, and scores lower on admitting uncertainty. The increases in evidence-based responses and assumption disclaimers from low to max effort are comparisons across effort levels. They do not support an inference that other metrics also change in step with reasoning effort.
The original post does not specify the predecessor model, the number of pairwise comparisons, confidence intervals for each result, or the specific effort levels used for the response-count figures. The methodology page also does not publish the questions or generation parameters. These results cannot establish whether reduced uncertainty disclosure leads to more errors, nor can they be directly generalized to other clients, parameters, or tasks.
The judging framework can be reproduced using the methodology page, but the original results cannot be fully retested: the 900 questions are private, and per-question answers and comparison records have not been published. The original post does not provide all comparison versions, reasoning parameters, or the number of pairwise comparisons. Anyone running an independent test should record the exact model ID and effort level, API and generation parameters, number of questions, scoring rules, and item-level results, and label the results as independently tested rather than placing them directly alongside AA's scores.
GPT-6.2 Sol