GPT-6 Astra's 99.9% ARC-AGI-3 Score Rejected by ARCPrize: Only 62.7% Under Standard Conditions

新智元 · wechat · 2026-09-11

OpenAI's GPT-6 Astra launch spiraled into AI's most surreal moment: Jensen Huang declared "AGI has arrived" after training on 100K Grace Blackwell GPUs, but ARCPrize refused to certify the claimed 99.9% ARC-AGI-3 score, which came from OpenAI's custom tool-enabled environment; under the standard setup Astra scored only 62.7% at $26K per run (vs. Claude Opus 5 at 30.2%). ARCPrize conceded Astra beats humans on action efficiency by 51.7% in 96% of tasks but won't call it AGI. Brockman framed his "welcome to the AGI era" as a personal view, and OpenAI's contracts with Microsoft require independent expert verification—making the AGI question a matter of IPO-era equity stakes. Bloomberg notes no agreed AGI definition exists across OpenAI, DeepMind, Anthropic, and ARCPrize.

Original post →

More from AGI Musings

AGI Musings channel →