A chatbot model learned to measure a box with a tape measure: Claude Fable 5.2 tops a real robot-arm test, while Gemini breaks a gripper

Can the same AI models that write code and answer questions take direct control of a real robot arm? Independent AI benchmarking firm Artificial Analysis posted its latest answer on Wednesday evening US time: Anthropic's Claude Fable 5.2, released earlier this week, scored 75 on the firm's AA-Robotics v0.1 benchmark, taking first place from OpenAI's GPT-6 Astra (69) and GPT-6.1 Sol (65).

First, the test. Artificial Analysis launched AA-Robotics v0.1 in mid-October as a "research preview." The setup is simple: each model gets a six-axis tabletop robot arm (five joints plus a gripper) and a short goal. The instructions are minimal: how to reach the arm through a simple web API, the arm's name, and how to submit an attempt. The model watches the table through four cameras and writes its own code to plan and send movements. The eight tasks are everyday chores: press a red button, put a toy crocodile inside a rectangle, pour a drink into a cup, pack four blocks into a box and put the lid on, complete a shape-matching board, write "hello world" with a marker, build the tallest foam-block tower, and measure a box's length, width and height to within 2 millimeters. Every attempt is zero-shot, meaning no demonstrations, and a full pass scores 100 with partial credit for progress.

The headline is the box. In Artificial Analysis's first published results, every model scored zero or only 17 out of 100 on that task. The firm says Fable 5.2 is the first model to earn a score there, and it did it purely from video: it set up the tape measure next to the box and used the tape's markings as a scale to estimate the box's dimensions from the camera images. In plain terms, it figured out the human trick of using something of known length as a reference.

The same day brought a less flattering result. Artificial Analysis said Google's Gemini 4 Carbon spent 7 hours and 33 minutes in the lab "bashing the arm around so much it broke a gripper," and that the firm gave it a retry on that task.

Context matters. When it launched the benchmark, Artificial Analysis said these models take about 30 minutes on average per tabletop task, compared with roughly 20 seconds for a person, about 100 times slower. Even the most cost-effective model cost about 10 times as much as a human paid the average US hourly wage, so the firm called them "far too slow and expensive for any kind of production usage." It is also one firm, eight tasks and one model of arm, so the rankings could shift with a different test. The models run inside coding-agent tools such as Claude Code and Codex, which is very different from the specialized control models used in factory robots.

So why pay attention? A short time ago, having a general-purpose chatbot model drive a physical arm through tasks like these would have sounded far-fetched. Artificial Analysis argues that future robots will likely rely on strong general-purpose AI "brains," which is why it is tracking how these models handle the physical world. Reading a tape measure off a video feed shows that models' grasp of real-world images and spatial reasoning is getting more practical.

What it means for you: a robot that folds your laundry or clears your table is still far off. These systems are slow, expensive and occasionally break the hardware. But the direction is clear: the same AI that chats with you and drafts your emails may eventually be able to look at your room and do simple physical tasks. As models get faster and cheaper, scores on tests like this are worth watching.

Original sources published: Artificial Analysis (AA-Robotics v0.1 launch post), Oct 15, 2026, 06:37 Beijing time / Oct 14, 2026, 6:37 PM EDT; Artificial Analysis (Claude Fable 5.2 results), Nov 12, 2026, 10:00 Beijing time / Nov 11, 2026, 9:00 PM EST; Artificial Analysis (Gemini 4 Carbon broken gripper), Nov 12, 2026, 12:05 Beijing time / Nov 11, 2026, 11:05 PM EST