Jensen Huang declares AGI has arrived (September 7) — the same model placed second on the intelligence index

The three lines

• Claim — Huang posted 'AGI has arrived' on September 7, citing GPT-6 Astra, trained on more than 100,000 Grace Blackwell NVLink72 units
• Evidence — a perfect 450 on Korea's college entrance exam (GPT-5.6 Sol 448.5, Claude Fable 5.1 447.5). But the composite intelligence index reads Astra 55, Fable 5.1 57
• Rebuttal — Gary Marcus: no definition, no basis, one or two of ten criteria met. And every executive who declared it benefits if the claim holds

Key questions

What exactly did Jensen Huang say

**"AGI has arrived."** Nvidia chief executive Jensen Huang posted the line on X on **September 7, 2026** and congratulated the OpenAI team. Three specifics came with it. **① Training scale** — Astra was trained on **"more than 100,000 Nvidia Grace Blackwell NVLink72"** units (see 「What Grace Blackwell NVL72 is」). **② Elapsed time** — **four years** from ChatGPT through o1 to Astra. **③ What comes next** — **"400,000 GPUs will be operational."** **The three statements are not the same kind of statement.** ① and ③ are facts Nvidia knows directly: how many of its own chips were sold and how many more are going in. ② is an observation. **But "AGI has arrived" is a verdict, and a verdict requires a standard.** Huang did not state one. OpenAI's leadership framed it the same way. Chief executive Sam Altman called Astra **"the best model in the world at computer use, professional work, science, coding and cybersecurity."** President Greg Brockman said it **"can do anything a human can do with a computer."** Astra itself was released on **September 3** (「GPT-6 Astra released on September 3」).

Is a perfect exam score evidence of AGI

**It is a first place on one test and a second place on another.** The figure Korean coverage leaned on is the **college entrance exam, or suneung** — the single national test that gates university admission in Korea, taken by roughly half a million students on one day each November. Astra scored **450 out of 450** across Korean, English, mathematics, Korean history and four elective subjects. It was **the only model to score full marks**; **GPT-5.6 Sol took 448.5** and **Claude Fable 5.1 took 447.5**. Reported task timings accompanied it: **5 minutes 27 seconds** to find pet-boarding options, **2 minutes 51 seconds** to search job listings. **The gaps deserve a second look.** Out of 450, second place trails by **1.5 points** and third by **2.5** — **0.33 percent and 0.56 percent.** The word 'perfect' carries more distance than the numbers do. And the ordering is not stable across tests. On the **Artificial Analysis Intelligence Index**, which folds many benchmarks into one score, **Astra scored 55 and Claude Fable 5.1 scored 57.** **The model being named AGI came second there.** How much any benchmark says about a model's real capability is itself an old problem (「What a coding benchmark is」).

What is the counter-argument

**That no standard was offered.** **Gary Marcus**, professor emeritus at New York University and a longstanding critic of the field, said Huang **provided neither a basis nor a definition**. Measured against the **ten AGI criteria** Marcus assembles from the research literature, he argues Astra satisfies **one or two**. Korean academics quoted in the same coverage called the declaration **premature** and **"a victory lap without grounds."** **The argument runs because AGI has no agreed definition.** Three distinct definitions are in common circulation and companies each use a different one (「What AGI means」). When the definition is contested, **"it has arrived" and "it is far off" can both be true of the same model.** There is a second point, about position. Huang runs the company that sells the chips the model was trained on, and in the same post referenced **the next 400,000**. Altman and Brockman run the company that built it. **All three occupy positions that benefit if the sentence is true.** That does not make it false. But a verdict rendered without external review is the same structural problem already visible in AI risk grading, where the company being graded does the grading (「What an AI risk grade is」·「What a preparedness framework is」). Astra is, after all, the model its own maker rated Critical for cyber capability days earlier (「OpenAI put a brake on its own model」).