OpenAI가 GPT-6 Astra를 위한 100,000개 GPU 훈련 실행을 완료했으며, 이는 프론티어 모델 훈련 규모에 있어 주요 이정표를 나타낸다.

OpenAI shipped GPT-6 Astra on September 3, 2026, with a detail that cut through the usual benchmark noise: the model was trained on more than 100,000 GPUs in a single run, according to OpenAI President Greg Brockman. Nvidia CEO Jensen Huang seized on the milestone to declare that "AGI has arrived." The claim sparked immediate pushback. Independent benchmark trackers showed Astra tied, rather than beat, Anthropic's Claude Fable 5.1 on several measures, while hardware analysts questioned whether Nvidia was announcing a genuine technical breakthrough or marketing its own GPUs. The reality is more nuanced: Astra represents a real infrastructure milestone, but one whose capabilities don't yet match the hype around it.

**Inside the 100,000-GPU Training Run**

Brockman detailed the scale in an interview with Ben Thompson on Stratechery on September 4: "This is the first run that we've trained on more than 100,000 GPUs, which is an easy number to throw around, but just think about the scale of that." The achievement wasn't simply running more chips in parallel; it was coordinating a single training job across six figures of accelerators rather than splitting work across smaller clusters.

Aidan Clark, OpenAI's VP of Research for training, explained the engineering shift: the data center networking, inference kernels, and model architecture were designed together for that scale from the start, not retrofitted afterward. Astra is also reportedly the first OpenAI model where earlier generations of its own models meaningfully contributed to the training pipeline itself—a departure from purely human-curated data toward models helping build the next generation.

This infrastructure jump is real. Coordinating more than 100,000 GPUs on one job requires networking, cooling, and fault-tolerance engineering that most AI labs hadn't publicly demonstrated before. But it doesn't automatically translate to broader capability leadership.

**Jensen Huang's "AGI Has Arrived" Moment**

Huang wasted no time attaching Nvidia's brand to the milestone. In a public post, he wrote: "GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. AGI has arrived." He framed the jump as rapid progress "from ChatGPT to o1 to Astra in 4 years" and noted that OpenAI already plans to bring 400,000 GPUs online for the next training run—a fourfold increase.

The hardware involved is Nvidia's Grace Blackwell NVLink72 racks, which pair Blackwell GPUs with Grace CPUs in a rack-scale system designed for coordinated multi-rack training. For Nvidia, tying "AGI has arrived" to its own hardware brand is a marketing win regardless of how the AGI debate resolves. Every headline about Astra amplifies headlines about the chips underneath it, a dynamic critics have flagged directly. TechRadar described Huang's declaration as reading more like a pitch for next-generation Nvidia GPUs than a neutral technical assessment.

**Where Astra Was Trained: Inside Stargate Abilene**

The training run happened at Stargate Abilene, OpenAI's flagship Stargate infrastructure campus in Texas, built by Oracle. According to Tom's Hardware, Oracle is deploying Nvidia's GB200 chips at the site, each combining two Blackwell B200 GPUs with a Grace CPU. Oracle plans to install more than 450,000 GB200 GPUs under a 15-year lease agreement. The site was still expanding through mid-2026, with additional buildings coming online.

Stargate itself is the broader $500 billion multi-year infrastructure program OpenAI announced with Oracle and SoftBank in January 2025, targeting roughly 10 gigawatts of AI computing capacity across multiple US sites. Abilene was the first campus to go live and is now the first site OpenAI has publicly credited with training a model on more than 100,000 GPUs at once. This wasn't a theoretical capacity plan—it was a real training job running on hardware that exists today, in Texas, under an active lease.

**GPT-6 Astra's Benchmark Scores, Explained**

OpenAI's system card reports Astra saturating several of its hardest internal evaluations. The model scores 99.9% on ARC-AGI-3, a test of abstract visual reasoning, and 97.6% on FrontierMath Tier 4, math problems designed to resist memorization. On ExploitBench, an internal measure of offensive cybersecurity capability, Astra scores 100%. On OSWorld 2.0, a benchmark for autonomous computer use, Astra scores 72.6% and completes tasks in roughly 40 minutes on average, compared with 65.7% and about 75 minutes per task for GPT-5.6 Sol.

Independent evaluators paint a more mixed picture. On the Artificial Analysis Intelligence Index, a third-party aggregate of reasoning, knowledge, and coding scores, Astra lands at 61.2, roughly level with GPT-5.6 Sol's 60.9 and behind Claude Fable 5.1's 65.7. On Deep SWE, a software engineering benchmark, Astra's result sits around 74.1%, ahead of Sol's 72.7% but within a point of a comparable Gemini Flash model's 73.8%. The pattern: Astra wins clearly on math and cyber capability, and holds roughly even on general coding, rather than leading across the board.

| Benchmark | GPT-6 Astra | Closest Comparison | Comparison Model |

|---|---|---|---|

| ARC-AGI-3 (abstract reasoning) | 99.9% | not yet published by rivals | — |

| FrontierMath Tier 4 (math) | 97.6% | not yet published by rivals | — |

| ExploitBench (offensive cyber) | 100% | not yet published by rivals | — |

| OSWorld 2.0 (computer use) | 72.6% in ~40 min/task | 65.7% in ~75 min/task | GPT-5.6 Sol |

| Deep SWE (coding) | ~74.1% | 73.8% | Gemini Flash-tier model |

| Terminal-Bench 4.0 | 57.7% | 55.8% | Claude Fable 5.1 |

| Artificial Analysis Intelligence Index | 61.2 | 65.7 (leads) | Claude Fable 5.1 |

**The First "Critical" Cyber Capability Rating**

Astra's benchmark sheet comes with a warning label. Under OpenAI's Preparedness Framework, Astra is the first OpenAI model to reach the "Critical" tier of cybersecurity capability, meaning that with the right tools and access, it can find previously unknown security flaws and develop new [article text cuts off in source]