OpenAI's GPT-6 Astra Tested on Robot Arm Tasks Compared to Fable Models

Sep 6, 2026

AI Summary

OpenAI's GPT-6 Astra was evaluated on two robotic tasks using YAM arms, showing strong performance in one task but struggling in another. In the block placement task, Astra succeeded in 19 out of 20 trials, while its performance on the puzzle task was less effective, completing only 2 out of 20 trials.

The evaluation involved OpenAI's GPT-6 Astra and two versions of the Fable model, Fable 5 and Fable 5.1, using YAM robotic arms under the same policy for two tasks: placing a red block in a bowl and inserting a blue puzzle piece into a groove.
In the block placement task, Astra successfully completed the task in 19 out of 20 trials, while Fable 5.1 completed 8 out of 20 and Fable 5 only 1 out of 20. Astra took an average of 2.5 minutes per trial, costing approximately $0.94 per run.
For the puzzle task, Astra completed the insertion only 2 times in 20 trials, similar to Fable 5.1, which also completed 2 out of 20. Astra's cost per run for this task was about $1.36, compared to Fable 5.1's $2.18.
Each trial was scored by a human grader based on the highest stage reached, with a total of 120 trials conducted across both tasks.

gpt-6robot armsai technologyautomationmachine learning

OpenAI's GPT-6 Astra Tested on Robot Arm Tasks Compared to Fable Models

Related Stories

Exploring the Role of Humanoid Robots in Addressing Elder Care Challenges

European Chemical Companies Compete in Humanoid Robotics Market Driven by PEEK Demand

Morgan Stanley Analyst Highlights Potential of Physical AI to Boost Global GDP

AstroForge Develops AI-Controlled Spacecraft for Asteroid Mining