Article 324: ARC Prize Scores Astra at 99.9% and 62.7%. Same Model. The 37-Point Gap Is the Harness
On 3 September, OpenAI launched GPT-6 Astra with 99.9% on ARC-AGI-3. Jensen Huang quoted the figure on X three days later while declaring "AGI has arrived." Greg Brockman closed the launch briefing with "Welcome to the AGI era." The number travelled everywhere.
The number is real and ARC Prize published it. It is also not the number any other model is measured by.
Two harnesses, one model
A harness is the scaffolding between a model and a benchmark: how prompts arrive, what state survives between requests, how much of its own working the model can carry forward. On an interactive benchmark like ARC-AGI-3, where an agent has to infer novel rules by playing unseen games, the harness is half the measurement.
- OpenAI's adapter: 99.9% on the Semi-Private set. It preserves the model's richer internal reasoning state between turns.
- ARC Prize's Standard harness: 62.7%. Same model, provider-neutral conditions. The model keeps only notes it explicitly chose to keep; its opaque reasoning state does not persist, so each request starts effectively cold.
62.7% is still the highest score ever recorded on ARC-AGI-3, and still far ahead of every competitor. The 99.9% is not fabricated - ARC Prize ran it, publishes it, and labels it by harness.
The correction that needs its own correction
A wave of digests this week framed the story as ARC Prize "rejecting" OpenAI's figure. That is not quite what happened, and the distinction matters. ARC Prize's objection is to ranking, not to the number: no rival model has ever been run under OpenAI's adapter, so a 99.9%-versus-70%-comparison is measuring two different things and calling it a leaderboard.
There is an unflattering reading in both directions. OpenAI's adapter is higher-scoring partly because it is also faster and cheaper - it does less re-derivation per turn. ARC Prize's standard rig is deliberately harder because persistence of private reasoning is exactly the capability it does not want to hand out for free. Neither side is lying. They are answering different questions under the same headline.
What to do with this
When the next 99%-style benchmark claim lands, ask three things before you repeat it: which harness, who ran it, and has any competitor run under the same one. If the answer to the third is "no," you have a demonstration, not a comparison.
Key Takeaway: The most useful thing to come out of this is not 62.7%. It is the demonstration that a 37-point swing needs no dishonesty at all - just a different definition of what the model is allowed to remember. That is now the single biggest unexamined variable in frontier benchmarking, and almost nobody reports it alongside the score.
Sources: ARC Prize results table for GPT-6 Astra; The Next Web, "OpenAI's AGI number came from a harness, not the model," September 2026; Toolpilotpro analysis, updated 26 October 2026; LLMBytes, 25 October 2026