GPT 6.2 两个新模型口碑遇冷:Astra 偏保守,Sol 稳但平淡
OpenAI 这次 GPT 6.2 的 Astra 和 Sol 被骂保守,但作者拆开对比了 Claude Fable 5.2,发现是过度优化惹的祸。想了解两个顶级模型到底差在哪、以及 LM Arena 之外的评测新玩法,这篇说得挺实在。
OpenAI 的 GPT 6.2 系列包含两个新模型:Astra 和 Sol。社区反馈整体偏负面,主要不满在于两者相较 Claude Fable 5.2 和 Claude Opus 5.5 这两款竞品模型没有拉开差距。从实测反馈看,Astra 6.2 输出干净、准确但风格过于克制简洁,Claude Fable 则更有创意、更具趣味性,代价是准确度稍差。
作者认为这次发布被批评得有些苛刻:两个模型相对自家前代确实有进步,只是没能胜过对手的搭档模型,Astra 甚至有部分能力被指略有退步。作者推测 OpenAI 受近期 Hugging Face 事件影响,对模型做了过度优化或刻意收敛,导致输出显得保守。对这次 .2 小版本更新而言,这种取向可以接受,后续大版本需要改善。
文章还提到 LM Arena 排行榜长期与真实体验脱节,作者引用了 Peter 的视频,展示了一种直接对比顶级模型输出的评测方式,并希望该对比界面能上线官网供用户试用。
It seems the GPT 6.2 releases have not gone down well, but I think that is quite harsh, and it needs to be put into perspective.
These two models, Astra and Sol, have improved upon their predecessors. What the community is moaning about is that they have not improved upon their companion models in Claude Fable 5.2 and Claude Opus 5.5, which is, I admit, disappointing. That is not to say they have not improved, although a few would argue Astra has degraded somewhat in part.
From what I can tell from watching and reading Thank you.others, certainly with the top models, Astra 6.2 is muted, boring, overly concise, but clean, and accurate, whereas Fable is more creative and interesting, but less accurate and cartoonist in places.
What seems to have happened is OpenAI have been hit by this Hugging Face incident and have toned down or over-optimised their model too much. I think that is fine for this .2 release, but it certainly should be improved upon in future models.
I've been very critical of LM Arena in the past for the leaderboards just not showing what we all know as the reality but I just watched Peter's video below and is a great way to compare the models.
I haven't checked but I hope this interface that Peter demonstrates is on their website because I would love to try that myself and see these models and the prompts that were used to generate them.
本条情报归属 AITOP「模型」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「中美大模型技术代差与商业化对照」「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:GPT-6.2、OpenAI、Claude、LM Arena,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「@koltregaskes」(@koltregaskes),经 AITOP 多维评分管线结构化提炼。