GPT‑6 Astra on robotic manipulation

September 4, 2026

A follow‑up to our comparison of Claude Fable 5 and Fable 5.1. We gave OpenAI's GPT‑6 Astra control of the same YAM arms under the same Inspect Robots agent policy, on the same two tasks:

“Pick up the red block from the table and place it inside the bowl.”

“Pick up the round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board.”

On the bowl task Astra placed the block in 19 of 20 trials, against Fable 5.1's 8 of 20 and Fable 5 in 1 of 20, in 2.5 minutes per trial to Fable 5.1's 6.8, at an estimated $0.94 per run to $2.12.

The puzzle task is a different story: Astra completed the insertion 2 times in 20 against Fable 5.1's 2 in 20. It reaches the groove and stalls at the same final step Fable does, at $1.36 per run to $2.18.

Block into bowl: the best completed run of each model (highest stage, then shortest), each played in its own time at the same speed‑up. Timers show real elapsed time with thinking pauses removed.

Astra completes the bowl task far more often, at about half the cost per run

Large dots are condition means; faint dots are individual trials (100 if completed, 0 otherwise) at their own cost.

Scoring

Every trial was scored by a human grader on the highest stage it reached, so a run that fails still records how far it got. The rubric is unchanged from the Fable report.

Astra places the block almost every time; on the puzzle it stalls where Fable does

Share of trials per model reaching each stage; n per row is the number of trials in that cell.

Results

All runs

Every counted trial, 120 in total.

Technical specifications

Limitations

• Astra's trials were run two days after the Fable trials, and not interleaved with them. The puzzle comparison is on the same rig; the bowl comparison is not: the Fable bowl trials ran on rig-3, which was unavailable.
• Grading was operator-judged with the model known, so scores are open to unconscious bias.
• Costs are list price. Anthropic requests were sent without prompt caching; OpenAI cached about a fifth of Astra's input automatically, which is not discounted here, so Astra's cost is, if anything, overstated.
• Objects were reset by hand between trials, and all models ran at medium reasoning effort only.