GPT-6 Astra's Korean CSAT 450 and Jensen Huang's AGI line: what's confirmed
Korea's 2026 CSAT LLM dashboard put GPT-6 Astra alone at 450/450. Jensen Huang then said AGI has arrived. Here is what Korean press actually confirms—and what still needs an asterisk.
TL;DR — On Korea's public 2026 CSAT LLM Solution Log, GPT-6 Astra was the only evaluated model to score a perfect 450/450 across Korean, English, math, Korean history, and four electives—with no web tools. Nvidia CEO Jensen Huang then wrote on X that "AGI has arrived." Both lines are real headlines. Neither is a government certificate that AGI exists, and the exam result is a GitHub dashboard score, not a claim that Astra sat a live CSAT.
Korean news this week is not mostly arguing about OpenAI's launch deck. It is arguing about a perfect score on the country's college-entrance exam benchmark—and about whether executives get to call that AGI.
What the CSAT dashboard actually shows
The Korea Times (from Hankook Ilbo) and AI Times both describe a public GitHub evaluation—the 2026 CSAT LLM Solution Log—updated around 6–7 September 2026. Models answer 2026 CSAT items in Korean language, English, mathematics, Korean history, plus four electives often treated as hard: Physics I, Chemistry I, Life Science I, and Society and Culture. Runs use the official model APIs. No search, no calculator, no custom system prompt.
On that board, GPT-6 Astra alone hit 450. Press rankings put GPT-5.6 next at 448.5, then GPT-5.4 at 448, Anthropic's Claude Fable 5.1 at 447.5, and Gemini 3.1 Pro at 445. Earlier "perfect" runs for Gemini 3.1 Pro and GPT-5.4 were framed as a student-style load (two electives). Astra is the first reported clean sweep on this wider four-elective layout.
Token thrift is part of the story. Korea Times cites about 357,000 tokens for Astra—the lowest among publicly listed peers in that write-up (GPT-5.6 ~429k; Fable 5.1 ~562k). AI Times quotes the dashboard maintainer, Soonchunhyang student Gu Yu-gyeom, on unusually low output tokens versus prior frontier models. Treat those as dashboard / reporter figures, not OpenAI marketing slides.
What executives said next
OpenAI launched GPT-6 Astra on 3 September 2026. President Greg Brockman had already floated "AGI era" language at launch (covered in our sister Terapep explainer on confirmed vs not). On 7 September, Money Today reports Nvidia CEO Jensen Huang posting that Astra was trained on 100,000+ Grace Blackwell NVLink72 systems, that the path from ChatGPT through o1 to Astra took four years, and that "AGI has arrived" with congratulations to OpenAI—plus a forward look at 400,000 GPUs.
That is a chipmaker CEO cheering a customer milestone on X. It is not NIST, not a Korean ministry, and not an independent scientific body certifying AGI.
What Korean coverage still hedges
Training contamination. CSAT items and answers were public. Korea Times notes you cannot rule out prior exposure—while also noting rival models sat the same open conditions.
Other scoreboards disagree. Money Today cites Artificial Analysis's AAII with Astra around 55 versus Claude Fable 5.1 around 57. Perfect CSAT pixels do not automatically mean "smartest on every chart."
Exam ≠ workplace. Korean AI voices quoted in Korea Times say open-ended work with no answer key matters more than OMR-sheet saturation.
Access drama is a parallel story. Foreign Plus subscribers this week are mostly arguing about finding Astra in Work / Codex rather than the ordinary Chat picker, messy rollout apologies, and how fast quotas burn. That is product UX and billing—not the CSAT dashboard.
| Claim | Status | Source |
|---|---|---|
| GPT-6 Astra scored 450/450 on the 2026 CSAT LLM Solution Log (4 electives, no web) | Confirmed as dashboard / press reporting | Korea Times; AI Times |
| First AI on that wider four-elective layout to hit a clean 450 | Reported by Korean tech press | AI Times; Korea Times |
| ~357k tokens; lowest among peers listed in Korea Times | Reported | Korea Times |
| Jensen Huang X post: "AGI has arrived" (7 Sep 2026) | Reported | Money Today |
| Astra formally certified as AGI by an independent standards body | Not confirmed | — |
| Astra sat a live, proctored CSAT as a human examinee | Not confirmed | Dashboard is an API eval |
| Astra leads every external intelligence index | Not confirmed (AAII reported behind Fable 5.1) | Money Today |
Why this story travels outside Korea
수능 is the cultural shorthand foreigners usually hear as "Korea's SAT, but harder and more decisive." A model clearing a public full-subject CSAT log is therefore an easy viral object—even when the method is an API harness on GitHub. Pair that with Huang's AGI line and you get two exportable soundbites. The useful English frame is narrower: what the dashboard measured, what executives claimed, and what still needs an asterisk.
FAQ
Did GPT-6 Astra "pass the real 수능"?
It topped a public LLM solution log built from 2026 CSAT materials. That is not the same sentence as "registered, sat in a classroom, filled an OMR sheet."
Did Nvidia declare AGI as settled science?
Huang's post, as relayed by Money Today, is congratulatory language. Keep it labeled as an executive quote.
Is this the same as OpenAI's launch "AGI era" line?
Related PR weather, different speaker and date. Brockman's launch framing is separate from Huang's 7 September post.
Sources
- GPT-6 Astra aces Korean college entrance exam (The Korea Times)
- GPT-6 아스트라, AI 최초 수능 완전 만점 (AI Times)
- "AGI가 왔다" 젠슨 황도 선언…수능 만점 받은 GPT-6 아스트라 (Money Today)
- GPT-6 Astra: A new generation of intelligence (OpenAI)
- Sister explainer (access, API pricing, cyber gating): GPT-6 Astra: what's confirmed and what isn't
Image: "OpenAI logo 2025 (symbol)" by OpenAI, public domain, via Wikimedia Commons. Brand mark only — not a screenshot of GPT-6 Astra or a CSAT paper.
← Back to all posts