Skip to content
Advertisement
GPT-6.2 Astra tops AA-Robotics v0.1 robot arm benchmark with a score of 79

GPT-6.2 Astra tops AA-Robotics v0.1 robot arm benchmark with a score of 79

OpenAI's latest model leads Artificial Analysis's real-world robotics test, but it still needs about 30 minutes to do what a person does in 20 seconds

OpenAI’s GPT-6.2 Astra now sits at the top of AA-Robotics v0.1, a benchmark that asks language models to control a real robot arm. It posted a score of 79 and completed six of eight physical tasks.

The leaderboard was published on November 17, 2026. The ranking matters because it measures models moving actual objects, not answering trivia or writing code in a sandbox.

What the benchmark actually tests

Artificial Analysis launched AA-Robotics v0.1 on October 14, 2026. The setup pairs frontier language models with a Hiwonder NexArm, a 6-axis robot arm, through a coding-agent interface.

The model watches through camera feeds and issues commands over HTTP APIs, then has to figure out whether its plan worked.

The eight tasks are deliberately ordinary tabletop chores. They include pressing buttons, pouring drinks, stacking blocks, writing “hello world,” and measuring objects to within 2 mm.

The leaderboard and the gains

GPT-6.2 Astra’s 79 put it narrowly ahead of GPT-6.2 Sol, which scored 78. Claude Fable 5.2 followed at 75.

Advertisement

The jump from earlier snapshots is the more interesting number. In October, GPT-6 Astra had scored 69.4, while GPT-6.1 Sol landed at 64.6 and Gemini 4 Argon trailed at 55.3.

The leaderboard also had no open-weights models on it.

Slow, pricey, and a little clumsy

The models needed approximately 30 minutes per task, compared with a human benchmark of about 20 seconds.

Cost adds another wrinkle. Per-task costs ranged from $1.93 to $16.36 depending on which model was running the arm.

Grasp reliability fluctuated between 28% and 45%.

No model initially completed the box-measurement task, the one demanding accuracy within 2 mm.

Why Artificial Analysis built this

AA-Robotics takes the coding-agent skills that labs have been sharpening and points them at hardware. The model is not trained as a robot controller. It writes code, watches what happens through a camera, and tries again.

Artificial Analysis frames the project as a way to encourage rapid iteration in embodied AI, the field focused on systems that act in the physical world. It plans to add more tasks and keep refining the evaluation.

What this means

For AI labs, the leaderboard turns physical competence into a visible scoreboard. OpenAI holds the top two spots, with Claude Fable 5.2 close behind, so the race is tight enough that the next release cycle could reshuffle the order.

For businesses weighing automation, the more relevant figures are the 30 minutes, the $1.93 to $16.36 per task, and the 28% to 45% grasp reliability.

The absence of open-weights models is also worth watching. If an open model eventually shows up on the board, it would signal that robot control via language models is spreading beyond a handful of frontier labs.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.