GPT-6 Astra on robot arms

OpenAI's GPT-6 Astra outperformed OpenAI's previous models Claude Fable 5 and Fable 5.1 in a series of tasks involving robot arms. In the "Pick up the red block from the table and place it inside the bowl" task, Astra successfully placed the block in 19 out of 20 trials, compared to Fable 5.1's 8 out of 20 and Fable 5's 1 out of 20.

The task was completed in 2.5 minutes per trial for Astra, compared to 6.8 minutes for Fable 5.1 and a time that was not specified for Fable 5. The cost per run for Astra was estimated at $0.94, while Fable 5.1 and Fable 5 cost $2.12 each.

The "Pick up the round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board" task proved to be more challenging for all three models. Astra managed to complete the puzzle only 2 times out of 20, with Fable 5.1 succeeding in 2 out of 20 trials. Both Astra and Fable 5.1 reached the final step of the puzzle, but were unable to complete it.

The cost per run for Astra in this task was estimated at $1.36, while Fable 5.1 and Fable 5 cost $2.18. The data for each model was collected through separate trials, played at the same speed-up. Timers recorded the real elapsed time, excluding thinking pauses.