GPT-6 Astra hits 66% on ARC-AGI-3, up from 8%, as ARC Prize plans AGI-4 around open-ended invention
GaryMarcus · x · 2026-09-04
ARC Prize's Mike Knoop announced that GPT-6 Astra is the new SOTA on ARC-AGI-3, scoring 66% as a direct model at $500/game — a qualitatively large leap from Sol's 8% on comparable verified scores. ARC v3 tests whether models can make sense of unfamiliar environments and autonomously pursue goals within them.
Knoop cautioned this still isn't evidence of AGI: open-ended invention and discovery remains undemonstrated by any model and will form the basis of ARC-AGI-4. Kenneth Stanley added that benchmarks keep rising while open-endedness grows in importance.
Related event: GPT-6 Astra Saturates ARC-AGI-3, Raising Benchmark Validity Questions(10 posts)→
More from AGI Musings
- Anthropic's Joshua Saxe: deep learning's core science questions are being abandoned — joshua_saxe · 2026-09-04
- Desktop app or CLI? 'Operating system' is the third answer for AI's future — majidmanzarpour · 2026-09-04
- Why I'm 0% Worried About AI Killing Everyone: The Case Against Apocalypse Thinking — granawkins · 2026-09-04
- AI commentator: kids shouldn't outsource cognitive sovereignty to AI — PolarBearby · 2026-09-04
- Jim Fan on World of Bits: OpenAI's 2016 booking-agent dream finally realized by GPT-6 — DrJimFan · 2026-09-04
- The bitterest lesson: your production job is their ablation — vedantmisra · 2026-09-04