GPT-6 Astra Launches with Benchmark Leaks: ARC-AGI-3 Hits 98.6%

OpenAI 于 09-04 发布 GPT-6，代号 Astra，基准测试成绩随发布同步流出，多项指标大幅刷新纪录。据称这是 OpenAI 迄今规模最大的训练运行，也是首个在 Stargate 基础设施上用超过 10 万 DBU 预训练的模型。

已确认

• 多条帖子一致转述的基准成绩：ARC-AGI-3 从 7.8% 升至 98.6%，FrontierMath Tier 4 v2 达 97.6%，DeepSWE v1.1 达 74.1%，BenchCAD 95.9%，GPQA Diamond 96%，ExploitBench 满分，在展示的所有测试中全面超过 GPT-5.6 Sol（m4、m5、m11、m14 等）
• 官方博客口径：AAII 得分 61.2，略高于 GPT-5.6 Sol；Coding Agent Index 得 67，略落后于 Fable 5（m6、m17）
• 训练细节：研究员 Aidan Clark 称这是 OpenAI「迄今规模最大的训练运行」，首个在 Stargate 基础设施上以超 10 万 DBU 预训练的模型，且首次由前代模型主要参与监督下一代模型的训练（m10、m15、m20）

尚未确认

• 多数成绩以截图形式流传，部分帖子（m8、m13）明确标注未经官方证实，个别帖子提醒发布前跑分爆料常有夸大或伪造；FrontierMath Tier 4 得分在不同帖子中分别出现 97.6% 与 97.4% 两种说法，具体数值有待官方口径确认

为什么重要

• ARC-AGI-3 从 7.8% 跃升至 98.6% 的幅度极为罕见，randlongevity 表示完全没预测到该分数能这么高，认为这印证了「我们正活在奇点里」（m19）
• bindureddy 实测后给出更细的判断：Astra 在推理、数学、数据分析和研究能力上确实胜过 Fable 5.1，但 Fable 5.1 仍是编程之王（m12），与官方编码指数略逊 Fable 5 的数据互相印证
• teortaxesTex 借此感叹模型迭代节奏过快：GPT-5.6 Sol 发布不到两个月即被超越，「享乐跑步机」已达峰值，刚发布就过时（m7）

2026-09-04 ~ 2026-09-04 · 31 related posts

Full story(2 episodes)→

• Episode 1: Astra Rumored to Have Two Test Builds Focused on Long-Horizon Autonomy(2026-09-03, 2 posts)
• Episode 2: GPT-6 Astra Launches with Benchmark Leaks: ARC-AGI-3 Hits 98.6%(2026-09-04, 31 posts)

• OpenAI's Astra trained on largest-ever run using 100K+ GPUs at Texas Stargate site — cedric_chee · 2026-09-04
• GPT-6 Astra benchmarks: how the new model scores — CounterReady4774 · 2026-09-04
• GPT-6 Astra Benchmarks Released — james_6732 · 2026-09-04
• Leaked GPT-6-Astra benchmarks claim 98.6% on ARC-AGI-3, unverified — scaling01 · 2026-09-04
• Unconfirmed GPT-6 Astra benchmark screenshots circulate online — haider1 · 2026-09-04
• GPT-6 Astra hits 97.4% on FrontierMath Tier 4, trained on 100k+ GPUs — ChrisGPT · 2026-09-04
• Unconfirmed GPT-6 Astra benchmarks leak amid 'AGI era' launch claims — mark_k · 2026-09-04
• GPT-6 Astra's SWE results show only ~6% gain over its predecessor, critic notes — robleclerc · 2026-09-04
• GPT-6 Astra beats Fable 5.1 on Terminal science and Automation benchmarks — ChrisGPT · 2026-09-04
• Leaked OpenAI Astra scores claim 98.6% on ARC-AGI-3 and near-perfect math benchmarks — teortaxesTex · 2026-09-04
• [source] GPT-6 Astra full benchmarks: ARC-AGI-3 jumps from 7.8% to 98.6%, GPQA 96% — kimmonismus · 2026-09-04
• ARC-AGI-3 hits 98.6% with GPT-6, settling a five-month-old saturation bet — burny_tech · 2026-09-04
• GPT-6 Astra reportedly scores 98.6% on ARC-AGI-3, months ahead of Chollet's timeline — burny_tech · 2026-09-04
• Experts stunned as OpenAI's Astra posts a shocking ARC benchmark result — salahuddin · 2026-09-04
• Astra Is Out and So Are the Benchmarks — Hyleal · 2026-09-04
• [source] GPT-6 Astra benchmarks leak: 98.6% on ARC-AGI-3, biggest training run yet — Dr_Singularity · 2026-09-04
• More GPT-6 Astra training details: first model supervised by prior models — Dr_Singularity · 2026-09-04
• OpenAI's Astra beats Fable 5.1 on reasoning and math, but Fable 5.1 remains the coding king — bindureddy · 2026-09-04
• GPT-6 Astra reportedly scores 61 on Artificial Analysis Intelligence Index, prompting questions — thesaraharminta · 2026-09-04
• GPT-6 Astra Early Benchmarks Look Meh, as Release Cadence Hits Peak Burnout — teortaxesTex · 2026-09-04
• GPT-6 benchmarks leak online ahead of official release — thesaraharminta · 2026-09-04
• OpenAI pre-trained Astra on 100,000+ GPUs, its largest training run ever — haider1 · 2026-09-04
• GPT-6 Astra reportedly scores 61.2 on AAII, slightly above GPT-5.6 Sol — Angaisb_ · 2026-09-04
• 'We're living in the singularity': researcher stunned by ARC AGI 3 score — rand_longevity · 2026-09-04
• Leak claims GPT-6 Astra scores 98.6% on ARC-AGI-3 and tops most benchmarks — yuwen_lu_ · 2026-09-04
• "If These Evals Are True, We Need Far More People Working on Evals" — HarveenChadha · 2026-09-04
• Exploit Bench at 100%? New Model's Cybersecurity Score Stuns Developers — jarrodwatts · 2026-09-04
• [source] GPT-6 Astra blog reveals 61.2 on AAII, Coding Agent Index of 67 slightly behind Fable 5 — Angaisb_ · 2026-09-04

3 near-duplicate retellings: teropa · iamaliveix · burhop