Leaked OpenAI Astra scores claim 98.6% on ARC-AGI-3 and near-perfect math benchmarks

teortaxesTex · x · 2026-09-04

A circulating report claims OpenAI's new model Astra scored 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD, 96% on GPQA Diamond, 100% on ExploitBench, and 98.6% on ARC-AGI-3.

If accurate, these would be unprecedented results across math, science, and software engineering benchmarks. However, the scores come from an unofficial forward and remain unconfirmed by OpenAI—treat as a credible-looking leak, not verified news.

Related event: GPT-6 Astra Launches with Benchmark Leaks: ARC-AGI-3 Hits 98.6%(41 posts)→

Original post →

More from Models

Models channel →