Leaked GPT-6-Astra benchmarks claim 98.6% on ARC-AGI-3, unverified

scaling01 · x · 2026-09-04

Account scaling01 posted purported benchmarks for a model called GPT-6-Astra: 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, and a perfect 100% on ExploitBench, after teasing a "nuke" hit on ARC-AGI-3. Unverified — no official source confirms these numbers, and some scores look implausibly high.

Related event: Leaked GPT-6 Astra benchmarks claim ARC-AGI-3 jump to 98.6%, unconfirmed by OpenAI(13 posts)→

Original post →

More from Models

Models channel →