ARC Prize retest puts GPT-6 Astra at 62.7% on ARC-AGI-3

On September 3, 2026, OpenAI launched GPT-6 Astra alongside a headline number: 99.9% on ARC-AGI-3, a benchmark built specifically to resist gaming, memorization, and brute-force scaling. Three days later, on September 6, Nvidia CEO Jensen Huang cited that figure on X when he declared that AGI had arrived, and the post drew more than 10 million views. OpenAI president Greg Brockman, speaking at the launch briefing, welcomed listeners to the AGI era.

Since then, the picture has been revised. ARC Prize released a re-evaluation of GPT-6 Astra on ARC-AGI-3 and reported two scores: the widely circulated 99.9% from launch, and a 62.7% measured under a provider-neutral Standard harness. According to ARC Prize, 62.7% remains the highest score ever recorded on the benchmark, but it sits 37 points below the headline figure, with the gap attributed to the harness rather than to the model itself. The reporting does not state when the re-test was run, so the discrepancy between the two numbers remains unresolved.