GPT-6 Astra scores 98.6% on ARC-AGI-3 in comparison table
A September 2026 release-notes comparison shows Astra far ahead on ARC-AGI-3, FrontierMath Tier 4 and ExploitBench, losing only slightly to Claude Fable 5 on HealthBench.
GPT-6 ASTRA SCORED 98.6% ON ARC-AGI-3, AND EVERY OTHER MODEL ON THE CHART ISN'T EVEN CLOSE. Side by side release notes, GPT-6 Astra next to Claude Fable 5.1 and Mythos 5.1, September 2026. The benchmark table: ARC-AGI-3, Astra 98.6% vs GPT-5.6 Sol 7.8%. FrontierMath Tier 4, Astra 97.6%, Claude Fable 5.1 and Fable 5 both at 87.8%. Agents' Last Exam, Astra 59.3%, Claude Opus 5 at 52.7%. AutomationBench, Astra 41.4% vs Claude Fable 5.1's 31.4%. Terminal-Bench Science, Astra 64.6% vs 52.6%. ExploitBench, Astra hits 100.0%, GPT-5.6 Sol trails at 78.5%. GPQA Diamond stays tight across the board, low-to-mid 90s for everyone. HealthBench Professional puts Claude Fable 5 slightly ahead of Astra, 60.9% to 63.4% margin closer than anywhere else on the sheet. One row at the bottom reads 0% for Astra, 0.29% for Sol, dashes for the rest, a benchmark hard enough that nothing on the chart has cracked it yet.
This post quotes another post: x.com/blazt_ai/status/2098493393932292515