GPT-6 Astra claims SOTA on ARC-AGI-3 at 66%, up from Sol's 8%
teortaxesTex · x · 2026-09-04
ARC benchmark founder Mike Knoop announced GPT-6 Astra as the new SOTA on ARC-AGI-3.
Key numbers
- Astra direct model scored 66% at roughly $500 per game
- A qualitative leap from Sol's 8%, apples-to-apples with all verified ARC v3 scores
What ARC v3 tests: whether models can figure out unfamiliar environments on their own and operate autonomously toward goals within them — Astra can now do this when properly equipped.
What's next: Knoop calls it a qualitatively large leap toward AGI but says evidence is insufficient to call it AGI. Human capability gaps remain, open-ended invention is unsolved, and this will form the basis of ARC-AGI-4.
Related event: OpenAI Launches GPT-6 Astra, Declaring 'Welcome to the AGI Era'(142 posts)→
More from Models
- Screenshots of Amazon Astra's Chain-of-Thought Output Circulate on Reddit — Tough_North7059 · 2026-09-04
- Theo: Astra is world-class at everything except frontends and mergeable code — adonis_singh · 2026-09-04
- Researcher: GPT-6 Astra turned a month-long research job into one week, feels RSI momentum — i_dg23 · 2026-09-04
- OpenAI says GPT-6 Astra shows Critical cyber capabilities, forcing harder ExploitBench evals — SIGKITTEN · 2026-09-04
- GPT-6 Astra coming to LMArena after zero-cherry-picking 3D world-gen gauntlet vs Claude Fable 5.1 — arena · 2026-09-04
- eyebench-v3 adds Muse-Spark 1.3 at 33%, Gemini-3.8-Flash flat vs predecessor — adonis_singh · 2026-09-04