GPT-6 Astra hits 95% on robot control, 2.3x cheaper than prior model

AaronBergman18 · x · 2026-09-06

A quoted post claims GPT-6 Astra scored 95% on a robot control task, up from 40% for Fable 5.1, while using 6.2x fewer output tokens at 2.3x lower cost. The poster's takeaway: robot control now runs in the same chat-transformer stack, with frontier models displacing specialized control architectures.

Related event: GPT-6 Astra scores 95% on robot control task at 43% of prior cost(5 posts)→

Original post →

More from Embodied

Embodied channel →