이 아티클은 아직 번역되지 않아 영어로 표시됩니다.
GPT-6 Astra Scored 99.9% on ARC-AGI-3. Is This AGI? The Footnote Explained
GPT-6 Astra scored 99.95% on ARC-AGI-3 with OpenAI's harness and 62.71% on the neutral one. What the gap means, the cost, and why it is not proof of AGI.
OpenAI released GPT-6 Astra on September 3, 2026. It writes code, uses a computer, runs research, and scored 99.9% on one of the most watched AGI benchmarks in the world. So the internet decided AGI is here.
There are two problems. Nobody agrees on what AGI means, and the 99.9 has a footnote.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
What ARC-AGI-3 tests
ARC-AGI-3 gives a model a puzzle it has never seen, with rules it must work out by playing. It is built so that memorising the internet does not help.
One model, two scores
ARC Prize ran Astra twice on two harnesses. A harness is the plumbing between the model and the test.
| Harness | Score | Compute cost |
|---|---|---|
| OpenAI's Provider Adapter (keeps hidden reasoning state alive between requests) | 99.95% | $18,817 |
| Neutral Standard harness | 62.71% | about $26,000 |
Same model, same benchmark. More money, lower score. The difference is the plumbing, and you will not get 99.9 from a normal API call.
ARC Prize says it directly: saturating the benchmark would not be proof of AGI.
Where it fails
ARC Prize published the per-game results for 25 games. On the neutral harness at maximum reasoning:
- 10 games: a flat 100%.
- Game G50T: 0%, at every reasoning level tested.
- BP35 and TU93: 2% and 3%.
That is not saturation. It is a model that is spectacular at most things and has a few blind spots it cannot see at all.
The number nobody screenshots
On 96% of the levels it played, Astra used fewer actions than the median human tested. Not faster, fewer moves. It worked out the rule with less poking at the board. A score tells you whether it got there. An action count tells you more about whether it understood.
What the score cost
ARC Prize also paid humans to take the test: $115 for a 90-minute session plus $5 per finished game, about $12.78 per game attempted. The headline run cost roughly 1,400 times what a person costs. ARC Prize notes that the human figure mostly measures willingness to sit an exam, not the energy cost of thinking.
The pattern across the three boards
| Benchmark | Astra |
|---|---|
| ARC-AGI-1 | 98.5% |
| ARC-AGI-2 | 95% |
| ARC-AGI-3 (neutral harness) | 62.71% |
Every time the test moves closer to doing something and away from knowing something, the score drops. That is the most useful thing on the page. Even so, other frontier models sit near the bottom of the ARC-AGI-3 chart, so 62% is still a real jump. It is smaller and more expensive than the headline.
Price and access
- $10 per million input tokens, $50 per million output tokens.
- A fast mode runs about twice as quick for about twice the price.
- Rolling out to a limited set of organisations first, then Plus, Pro, Business and Enterprise, plus the API, Azure and Bedrock.
The security rating
Astra is the first model OpenAI has rated Critical for cybersecurity under its own Preparedness Framework. In that framework, Critical means it can find and exploit unknown vulnerabilities in well-defended systems without a human guiding each step. The framework says that rating should trigger safeguards before release, not after.
So what is AGI?
Human-level intelligence. Most economically valuable work. General problem solving. An autonomous worker. The ability to learn any domain. Researchers use all of these, and they do not agree.
- If your definition is "scores well on hard tests", Astra qualifies.
- If your definition is "does a full day of someone's job alone without breaking anything", it is not close.
Benchmarks are clean. The real world has expired permissions, odd dialogs, bad Wi-Fi and CAPTCHAs.
Verdict
Did OpenAI build AGI? No. More exactly: there is not enough evidence, and the single number used to argue it falls to 62% when you change the plumbing.
Astra still matters. Coding, science, reasoning, computer use and tools are merging into one system. That is the move from AI you talk to, to AI you assign work to. A model that is unreliable at general intelligence can still be very reliable at a narrow task with tools, a loop and tests. That is an engineering change, and those are the ones that change next Tuesday.
Links
- Full video: https://youtu.be/CM0p_BovClk
- ARC Prize analysis: https://arcprize.org/blog/astra
- ARC Prize verified scores: https://arcprize.org/results/openai-gpt-6
- OpenAI announcement: https://openai.com/index/gpt-6-astra/
- GPT-6 Astra vs Claude Fable 5.1 for coding
What evidence would convince you that an AI is truly general?


