GPT-6 Astra on robot arms

OpenAI's latest AI model, GPT-6 Astra, was tested on the same robot arms used in a comparison of Claude Fable 5 and Fable 5.1. The two tasks evaluated were "Pick up the red block from the table and place it inside the bowl" and "Pick up the round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board."

In the bowl task, GPT-6 Astra successfully placed the block in 19 out of 20 trials, outperforming both Fable 5.1 (8 out of 20) and Fable 5 (1 out of 20). It achieved this in 2.5 minutes per trial, compared to Fable 5.1's 6.8 minutes. The estimated cost per run was $0.94 for GPT-6 Astra, $2.12 for Fable 5.1, and $2.18 for Fable 5.

The puzzle task, however, presented a different outcome. GPT-6 Astra managed to complete the insertion 2 times out of 20, matching Fable 5.1's performance. Like the bowl task, it reached the groove but stalled at the final step, costing $1.36 per run to $2.18.

The best completed run of each model was analyzed, with each trial played independently at the same speed-up. Timers recorded the real elapsed time, excluding thinking pauses. Individual trials were represented by dots (100 for completed trials, 0 for failed trials) at their respective costs. Every trial was graded by a human evaluator based on the highest stage reached, even if the run failed, the progress made was documented. The grading rubric remained consistent with the earlier Fable report.