GPT-6 Astra tops robot-arm benchmark

Astra looks like a meaningful jump in vision-language control, but this is a task-specific win—not evidence of general-purpose dexterity.

• –Bowl success reached 95%, versus Fable 5.1’s 40%, while Astra was faster at 2.5 minutes per run versus 6.8.
• –Astra also used fewer output tokens and cost roughly $0.94 per bowl run, suggesting a practical efficiency advantage for robotics experimentation.
• –On the harder puzzle task, Astra tied Fable 5.1 at 10% and repeatedly stalled at the final insertion step.
• –Different hardware rigs and non-interleaved bowl trials weaken the comparison, so the result should be treated as promising rather than definitive.
• –For developers, the strongest signal is the combination of a frontier model and an open agent harness—not autonomous, production-ready dexterity.