ModelsMulti-source

传 GPT-6.2 Astra 已达到千禧年大奖难题水平,OpenAI 内部模型更强

Why we picked it

有人根据泄露图表推测 OpenAI 内部 GPT-6.2 Astra 已经摸到千禧年大奖难题的水平,困惑度曲线显示预训练还在加速。想了解 OpenAI 模型迭代节奏和 GPT-5 之后进展的朋友可以看看这条分析。

一条推特分析称,解决千禧年大奖难题的模型很可能是 GPT-6.2 Astra,图表中该难题的门槛线恰好落在 Astra 的能力水平上。据称该结果来自 OpenAI 9 月初的内部模型,不到 3 个月后内部已有更新的模型。

Body

技术细节上,OpenAI 内部模型的困惑度比 Astra 还低约 0.03,大致相当于从 GPT-5 到 GPT-5.4 的提升幅度。OpenAI 研究负责人 Pachocki 表示,越接近困惑度下限,每一点困惑度下降对应的智能跃升越大。

这条推文还判断预训练进展仍在加速,OpenAI 官方也称 2026 年将出现陡峭的能力加速曲线。目前所有信息均为非官方推测,OpenAI 未发布正式公告。

Original · kimmonismus

1) The model that solved the Millennium Prize problem was most likely GPT-6.2 Astra. The chart puts the Millennium Prize line right at its level

2) OpenAI's internal model is already another ~0.03 lower in perplexity. That's roughly the jump from GPT-5 to GPT-5.4, and Pachocki says smaller perplexity gains mean bigger leaps in intelligence the closer you get to the floor

3) The Millennium Prize result came from an internal model in early September. Less than 3 months later, there's already a new internal model further down the curve

4) Looks like pretraining progress is still accelerating. OpenAI itself describes a steep acceleration in 2026

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 Astra, OpenAI, GPT-5, pretraining.

Source attribution: curated from kimmonismus (@kimmonismus). Cross-source corroboration: 11 independent sources have covered this development (including Greg Brockman, rohanpaul_ai, IT之家), confirming industry-wide signal strength.