Gemini 4 Carbon broke a robot gripper after 7.5 hours of flailing; the robotics test winner was Fable 5.2, which read a tape measure off the video

Model evaluator Artificial Analysis said on Nov. 12 that Google's Gemini 4 Carbon broke a robot gripper during its new test that hands a real robot arm to AI models.

"Gemini 4 Carbon broke one of our arms," the firm posted, saying Carbon spent 7 hours and 33 minutes in the AA-Robotics v0.1 lab bashing the arm around until a gripper broke, and that it was given a retry on that task. The post drew more than 500 likes.

Claude Fable 5.2 leads the benchmark with 75, ahead of GPT-6 Astra (69) and GPT-6.1 Sol (65). It is the first model to score on the measure-the-box task, which it did purely from video, leaving the tape measure in place and using its markings as a scale. Tests like this are spreading as AI models move from screens to physical machines.

Reactions split. "Every robotics lab has a guy who broke a gripper on day one, Gemini is just him now," one joked; another said to "send Google the repair bill." Others noted the tape-measure trick breaks when the markings are out of frame and called for a run with no scale visible, and asked whether Carbon's retry counted toward its score, which the firm has not made clear.