# GPT-6.2 Astra Scores 79 on Robot-Arm Benchmark With 668 Commands

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/62299
Published: 2026-11-20T20:00:24.000Z
Updated: 2026-11-20T20:00:24.000Z
Section: Technology

> Astra completed six of eight tasks, one point ahead of GPT-6.2 Sol. It used 92 commands for a box-cleanup task, about 75% fewer than Claude Fable 5.2.

GPT-6.2 Astra averaged 79 out of 100 on AA-Robotics v0.1, finishing one point ahead of GPT-6.2 Sol in an evaluation of models writing instructions for a physical robot arm.

Astra completed six of the benchmark’s eight tasks and issued 668 arm commands across them. GPT-6.2 Sol scored 78, while Claude Fable 5.2 scored 75.

On a box-cleanup task, Astra used 92 commands, about 75% fewer than Claude Fable 5.2. The benchmark places language models in a coding-agent role: They write instructions that a robot arm executes.

The evaluation uses a six-axis Hiwonder NexArm. Tasks include pouring drinks and stacking blocks, which call for spatial awareness and problem-solving. The models took about 30 minutes per task, compared with about 20 seconds for a person doing comparable work. Costs ranged from $2 to $16 per task, depending on the model and job.

AA-Robotics v0.1 is an early benchmark, and Astra’s lead over Sol was narrow. The results cover eight tasks.


