ARC Prize has rerun GPT-6 Astra on ARC-AGI-3 using its own standard harness and recorded 62.7%. OpenAI published 99.9% when it launched the model on 3 September 2026, and Nvidia chief executive Jensen Huang cited that number on X three days later to declare that AGI had arrived. The foundation called the 99.9% figure "not valid".
The gap is explained by the harness rather than the model. OpenAI's figure came from its own tools-enabled configuration, which ARC Prize says cost nearly $19,000 to run and under which no other lab's models are evaluated. Run the same model the way every other entrant is run and it scores 62.7% - which is still the leading result on the benchmark. Lower-compute Astra variants came in at 59.3% and 54.8%. ARC Prize also said Astra's progress does not qualify as AGI.
Two things make this worth more than a scoring argument. The first is that OpenAI reportedly revised five published benchmark figures after its launch post, which turns one disputed number into a pattern. The second is that the same week produced an independent study finding that harness choice alone moves coding-agent cost by up to 3x with no consistent accuracy gain. Both point at the same conclusion: a benchmark number without its harness is not a result, and most published numbers travel without one.
Sources
Startup Fortune