Astra beats itself on ARC-AGI-3 with 46% cost cut at max reasoning settings

daniel_mac8 · x · 2026-09-13

danielmac8 reports that Astra on max reasoning settings outperforms Astra on low reasoning settings on ARC-AGI-3 (standard harness) while costing 46% less. His hypothesis ties this to the "recurrent depth" reasoning mechanism reported by The Information before launch: the model can reason more overall without emitting a CoT token per step. He calls it a notable efficiency gain for complex tasks.

Original post →

More from Infra

Infra channel →