Reddit
RoboCurve put GPT-6 Astra on real robot arms: 19/20 on block-into-bowl versus Fable 5.1's 8/20, at 6x fewer output tokens
RoboCurve ran 20 trials per model per task on bimanual I2RT YAM arms (6-DoF, parallel-jaw grippers) through Inspect Robots 0.58.0 and published results on 2026-09-04. Astra hit 19/20 on block-into-bowl against Claude Fable 5.1's 8/20 and Fable 5's 1/20, averaging 2.1k output tokens versus Fable 5.1's 12.9k, $0.94 versus $2.12 per run, and 2.5 minutes versus 6.8. The counterweight is the second task: on puzzle-piece-into-groove both Astra and Fable 5.1 managed 2/20, so precision manipulation is untouched and the headline number covers only the coarse task.
↳ Follow the thread