GPT-6 Astra Outperforms Claude Fable 5.1 at Robot Pick-and-Place
In Robocurve's test, GPT-6 Astra put a block into a bowl in 19 of 20 trials versus 8 for Claude Fable 5.1, at less than half the cost. On a harder puzzle task, both managed 2 of 20.
TL;DR
- In Robocurve's robot-arm tests, OpenAI's GPT-6 Astra placed a block in a bowl in 19 of 20 trials, against 8 for Claude Fable 5.1 [^1].
- On a harder task, inserting a puzzle piece into a groove, both models succeeded 2 times in 20 and stalled at the same final step [^1].
Background
General-purpose AI models can now steer robot arms without robot-specific training by issuing commands through an agent harness. Robocurve had previously compared Anthropic's Claude Fable 5 and Fable 5.1 on the same setup; this report adds GPT-6 Astra [^1].
What happened
In a report dated September 4, Achu Menon and colleagues ran each model on a pair of YAM arms under the same Inspect Robots agent policy, with two tasks: pick up a red block and place it in a bowl, and pick up a round puzzle piece by its knob and place it into the matching groove [^1]. A human grader scored each trial by the highest stage it reached, from no purposeful approach to final placement [^1].
On the bowl task, GPT-6 Astra succeeded in 19 of 20 trials, against 8 for Fable 5.1 and 1 for Fable 5. It took 2.5 minutes per trial versus 6.8 for Fable 5.1, at an estimated $0.94 per run versus $2.12, and used about 2,100 output tokens per run against 12,900 [^1]. On the puzzle task, Astra and Fable 5.1 each completed the insertion twice; Astra reached the groove and stalled at the same final step as Fable, at $1.36 per run versus $2.18 [^1].
Why it matters
The results show a clear gap on a simple manipulation task and none on a precise one. Getting an object roughly to the right place is now reliable for the best general models; the fine, contact-heavy final step is where both still fail. The cost and token differences also matter for robots, where every extra reasoning step adds time between movements [^1].
Practical example
Say a lab wants a general model to sort parts into bins with a robot arm. Based on these trials, GPT-6 Astra would handle the drop-into-bowl type of task almost every time, faster and cheaper than Fable 5.1 [^1]. If the job instead requires fitting a part into a tight slot, neither model would be reliable yet.