Language models now drive a robot arm directly: top score 69 on eight tabletop tasks, about 30 minutes and $16 each

Frontier language models in coding agent harnesses can now directly manipulate a robot arm for simple tabletop tasks, but they remain far too slow and expensive for production use. A common comparison for language models directly controlling a robot arm was missing; AA-Robotics v0.1 reports zero-shot scores for five models on the same eight tasks. Five models ran the same eight tasks: GPT-6 Astra scored 69.4, GPT-6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8 and Claude Fable 5.1 46.3, with each attempt scored out of 100, full completion worth 100 and partial results earning points for progress.

Models plan and send arm movements by writing code rather than being trained as robot policies. Each model runs in an isolated container with brief instructions: how to reach the arm through a simple HTTP API, which arm it is (a Hiwonder NexArm, with no further specifications), and how to submit an attempt as complete. Models view the scene through four cameras, three around the table and one on the wrist, and can write and run code in the container to plan and send movements; after submission the end state is graded against the task's pass criteria, and all attempts are zero-shot.

Time and spend vary widely, and the averages hide large differences between tasks. In aggregate four of the five models look similar at about 30 minutes per task, while Gemini 4 Argon is slowest at 52 minutes; a human averages 20 seconds. GPT-6 Astra pressed the red button in 2.2 minutes, whereas Claude Fable 5.1 spent about ten minutes understanding and planning before its first action; on the crocodile placement task GPT-6.1 Sol declared itself complete after five minutes against 45 for Claude Opus 5.5, a ninefold difference; Gemini 4 Argon averaged 21 minutes observing and reasoning before its first action, nearly twice as long as any other model.

This evaluation is a research preview covering one 6-axis robot arm (5 joints and a gripper) and eight tabletop tasks spanning control and dexterity, spatial reasoning and adaptability; models run in an isolated container, reach the arm through an HTTP API, view the scene through four cameras, and all attempts are zero-shot. Human cost uses $37.81 per hour, the average hourly earnings of all employees on US private nonfarm payrolls (US Bureau of Labor Statistics, preliminary data for September 2026).

The release states that frontier language models are today far too slow and expensive for any kind of production robotics use, nearly two orders of magnitude worse than human baselines, and it expects this to change quickly. Models often continue refining attempts long after reaching a passing state; GPT-6.1 Sol builds a passing tower after about 17 minutes only to subsequently fail the task, and how models will improve their ability to reason about task completion in physical space remains to be seen. Gemini 4 Argon's activity times are estimated because Antigravity does not timestamp each step, so steps are placed between the exactly timed arm commands in proportion to their tokens.