Artificial Analysis released AA-Robotics v0.1 on October 14, 2026, a benchmark evaluating how well eight language models can understand, plan and perform physical tasks without prior training. Each model controls a six-axis Hiwonder NexArm robot, observes its environment through four cameras and receives a single objective through a coding-agent interface. OpenAI’s GPT-6 Astra ranks first with 69.4 points out of 100, followed by GPT-6.1 Sol with 64.6 points and Google’s Gemini 4 Argon with 55.3 points. To write “hello world”, GPT-6 Astra completed 514 arm movements in 13 minutes, compared with 3,270 movements and 34 minutes for Anthropic’s Claude Opus 5.5. Humans complete these tasks in about 20 seconds on average, while the models take approximately 30 minutes, making them roughly 100 times slower and substantially more expensive than human labor for the same work.
Source: Read the original article
