OpenAI’s GPT-6 Astra beats Claude Fable 5.1 on robotic bowl task
A comparison using YAM robotic arms found Astra succeeded in 19 of 20 bowl-placement trials, while both models managed the puzzle task twice.
OpenAI’s GPT-6 Astra reportedly outperformed Anthropic’s Claude Fable 5.1 in a robotic block-placement test, succeeding in 19 of 20 trials compared with 8 of 20 for Fable 5.1.
The comparison, published on an OpenAI-branded site and linked through Hacker News, gave both models control of the same YAM robotic arms under the Inspect Robots agent policy. They were tested on two tasks: placing a red block in a bowl and inserting a blue puzzle piece into a matching groove.
Astra was also reported to be faster and cheaper on the bowl task. Its average run took 2.5 minutes and cost an estimated US$0.94, compared with 6.8 minutes and US$2.12 for Fable 5.1.
The result was different on the puzzle task. Both models completed the insertion in two of 20 trials, with the source saying failures commonly occurred at the final insertion step. Astra’s estimated cost was US$1.36 per run, compared with US$2.18 for Fable 5.1.
The trials were assessed by human graders, and the comparison reported about 80 per cent fewer output tokens for Astra. The figures are estimates under the stated test conditions, and the supplied material does not include independent verification or enough methodological detail to establish broader robotic performance.
