ModelsMulti-source

GPT-6.2 Astra 与 Sol 风格大改:回复更短、更直接、更少认错

Why we picked it

OpenAI 这次改的不是能力是性格:回复砍了四成字数,先给答案再解释,结构化输出基本没了。做提示词的和做下游应用的要重新调格式依赖,而且模型更少承认不知道,建议在高风险场景里配事实核查。

Artificial Analysis 对 OpenAI 最新 GPT-6.2 两个变体 Astra 与 Sol 的行为追踪显示,与前代相比两者的对话风格出现系统性变化。Astra 平均回复长度从 352 词缩到 210 词,Sol 从 337 词缩到 223 词;Astra 的 Organization Index 从 70 降到 11,Sol 从 86 降到 27,说明模型不再依赖标题、列表等结构化输出。

Body

取而代之的是 answer-first 风格:Astra 的 Answer First 指标从 49 升至 89,Sol 从 64 升至 87,即先给结论再展开。同时 Natural Sounding 分数明显上升,Astra 达到 100,Sol 达到 79。不过两个变体承认不确定性的频率都大幅下降,Astra 从 96 跌到 24,Sol 从 98 跌到 51,这可能增加模型在不确定问题上显得过度自信的风险。

此外在高推理强度下 Evidence-Based 指标明显走高,Astra 从 12 升到 54,Sol 从 16 升到 68,表明高 effort 模式下模型给出更多可验证依据。这类风格指标的变化会直接影响依赖旧版输出格式的下游产品和提示词设计。

Original · Artificial Analysis

The personalities of GPT-6.2 Astra and GPT-6.2 Sol changed noticeably from prior releases: both use less structure, are less likely to admit uncertainty, and give shorter responses

➤ Shorter responses: 352 to 210 words per response (Astra), 337 to 223 (Sol) ➤ Less structured but more answer-first: Organization Index 70 to 11 (Astra) and 86 to 27 (Sol), while Answer First rises 49 to 89 (Astra) and 64 to 87 (Sol) ➤ Less likely to admit uncertainty: Admits Uncertainty 96 to 24 (Astra) and 98 to 51 (Sol) ➤ More natural sounding: Natural Sounding 58 to 100 (Astra) and 54 to 79 (Sol) ➤ More evidence-based at higher effort: Evidence-Based rises from low to max effort, 12 to 54 (Astra) and 16 to 68 (Sol)

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include GPT-6.2, OpenAI, Astra, Sol.

Source attribution: curated from Artificial Analysis (@ArtificialAnlys). Cross-source corroboration: 11 independent sources have covered this development (including The Rundown AI, Decoder, marktechpost), confirming industry-wide signal strength.