Jensen Huang: 'AGI Has Arrived' With GPT-6 Astra

Highlights

Nvidia CEO Jensen Huang said "AGI has arrived" on Sept. 6, crediting OpenAI's GPT-6 Astra, which was trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems.
Huang said another 400,000 GPUs are coming online next, underscoring the scale of ongoing AI infrastructure buildout.
Astra's marquee 99.9% score on the ARC-AGI-3 benchmark relies on a stateful evaluation harness; under a standard scaffold, the same model scores 62.7%.
OpenAI's own president, Greg Brockman, has said there is "no clearly defined AGI moment," undercutting the certainty of Huang's declaration.
API pricing for Astra starts at $10 per million input tokens and $50 per million output tokens, roughly 2.5x pricier than its predecessor.

Huang's AGI Declaration

Nvidia CEO Jensen Huang declared on September 6 that "AGI has arrived," pointing to OpenAI's newly launched GPT-6 Astra model as proof. In a post responding to a follower, Huang said Astra was trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems and credited the jump from ChatGPT to OpenAI's o1 reasoning model to Astra as happening in just four years. He added that another 400,000 GPUs are coming online next, a figure that speaks to the scale of compute Nvidia's hardware is now absorbing across the industry. OpenAI has since updated Astra's own product page with the performance, pricing and safety data behind the launch.

What the Benchmarks Actually Show

The benchmark numbers are genuinely striking on their face. Astra posted a 97.6% score on FrontierMath Tier 4, a test suite designed to be difficult even for top human mathematicians, and a 99.9% score on ARC-AGI-3, a benchmark built specifically to resist pattern-matching and reward genuine reasoning. On Terminal-Bench 4.0, Astra scored roughly 57.9%, ahead of OpenAI's own prior model GPT-5.6 Sol at 37.3% and modestly ahead of Anthropic's Claude Fable 5.1 at 55.8%; on the Agents' Last Exam benchmark, Astra's 59.3% edged out Claude Opus 5's 55.5%.

The 99.9% ARC-AGI-3 figure comes with an important asterisk, however: it was recorded under a stateful "provider adapter" harness that lets the model retain reasoning context and compact it between turns. Under the standard, stateless scaffold that ARC Prize uses to score most models, Astra's result drops to 62.7% — still strong, but far short of the headline number driving Huang's declaration. API pricing for Astra starts at $10 per million input tokens and $50 per million output tokens, roughly 2.5 times pricier than its Sol predecessor and in the same range as Fable 5.1.

Marketing Ahead of the Science

Huang's framing ran well ahead of OpenAI's own. Company president Greg Brockman has said there is "no clearly defined AGI moment," describing the transition to more capable systems as "more gradual than expected" rather than a single threshold being crossed. AI researcher Gary Marcus separately pushed back on Huang's claim, arguing it lacks a rigorous definition of AGI or independent evidence to support it. That gap between a hardware vendor's marketing framing and a model developer's own hedged language is itself the more interesting story: Nvidia has a direct financial interest in every new "AGI moment" claim, since each one reinforces the case for continued GPU capex at a scale that already runs into the hundreds of thousands of chips per training run.

For markets, the more concrete signal is the compute commitment itself. A single model trained on over 100,000 Grace Blackwell systems, with 400,000 more GPUs already lined up, is the kind of capital intensity that keeps Nvidia's earnings under scrutiny even at historically low valuation multiples, and it adds to a buildout wave that has Anthropic committing $45 billion to its own data center capacity ahead of a prospective IPO. The scramble extends beyond the US too — Washington has moved to rally domestic chipmakers in response to Huawei's competing bid for an AI data center contract in Egypt, another sign of how central this compute race has become to broader tech and trade policy.

What to Watch Next

The more measurable test than the AGI label itself is what happens to independent benchmark results once outside researchers get broader access to Astra beyond OpenAI's own reporting — specifically whether third-party evaluations under standard, stateless conditions converge closer to Astra's 62.7% ARC-AGI-3 score or its 99.9% headline figure. Also worth tracking is whether the 400,000 additional GPUs Huang referenced show up in Nvidia's next earnings disclosures as delivered shipments, which would validate the scale of buildout being described, or whether the timeline slips as it has for other announced compute commitments this cycle. Either outcome will matter more to how this moment is remembered than any single benchmark screenshot.

FAQ

What did Jensen Huang say about AGI?
On Sept. 6, Nvidia CEO Jensen Huang said "AGI has arrived," crediting OpenAI's GPT-6 Astra, which he said was trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems.

Why is Astra's 99.9% ARC-AGI-3 score contested?
That score was recorded under a stateful evaluation harness that lets the model retain context between turns; under the standard, stateless scaffold most models are scored with, Astra's result drops to 62.7%.

Does OpenAI agree that AGI has arrived?
Not exactly — OpenAI president Greg Brockman has said there is "no clearly defined AGI moment" and described progress as more gradual than expected, a more hedged framing than Huang's declaration.

How much does GPT-6 Astra cost to use?
API pricing starts at $10 per million input tokens and $50 per million output tokens, about 2.5 times pricier than its predecessor, GPT-5.6 Sol.