Same robotic arm writes hello world: one used 3,270 points, the other drew 514 strokes

The same robotic arm writing “hello world” produced two completely different processes from the two models. Artificial Analysis’s AA-Robotics v0.1 let the models control the robotic arm from their own perspective. On the left, Claude Opus 5.5 broke the letters into points and placed them one by one, using 3,270 robotic-arm movements and taking 34 minutes. On the right, GPT-6 Astra used continuous strokes, finishing in 13 minutes with 514 movements. The “hello” segment shown in the video has more detailed numbers: on the left, 2,491 movements and 30 minutes 35 seconds; on the right, 279 movements and 10 minutes 59 seconds. The full task and the segment are different, so the numbers change, but the direction of the gap does not.

First, to clarify the background. AA-Robotics v0.1 is a robotics benchmark created by Artificial Analysis. This comparison is about how models translate “writing hello world” into robotic-arm actions; chat responses and code completion are not part of this task. With the same robotic arm, the same task, and the same red pen, the models output different action sequences, and the resulting handwriting and timing are also different. The original article has more examples: AA-Robotics.

3,270 versus 514 is about 6 times as many movements; 34 minutes versus 13 minutes is about 2.6 times as much time. Some online commenters noted that each movement is not a teleport but a round trip, so writing point by point compounds the time cost. The explanation is intuitive: dot-matrix-style writing makes the robotic arm move up and down frequently, while continuous strokes reduce round trips. Others have compared this to tool calls in agents: if something one script can do is split into dozens of small calls, the number of actions balloons, and latency and cost balloon with it.

But the number of movements is not the whole story. A commenter calculated the rates: Opus made roughly 96 robotic-arm movements per minute, while Astra made roughly 40 per minute. The point-by-point model had a higher action frequency yet still finished later because it performed about 6 times as many actions. This detail is more interesting than simply “who is faster or slower.” Behind speed is action granularity, and behind granularity is how the model plans a continuous trajectory.

The red handwriting in the video also prompted a lot of off-topic comments. Some said it looked like writing on a wall with a red pen in a horror movie, and others asked why a red pen had to be used. This joking commentary does not affect the conclusion, but it shows that the presentation affects viewers’ first reactions. What is more worth watching in the experiment itself is the process: the left side looks like a dot-matrix printer, while the right side looks like a pen tip moving continuously along a line. One online commenter said Claude is rendering letters while GPT is writing. The claim may not be rigorous, but it precisely describes the impression in the video.

Another easily overlooked issue is error correction. A high number of movements could mean the model is planning point by point, or it could mean it is correcting bad placements while writing. From this video alone, it is impossible to tell how many of the 3,270 movements were corrections and how many were part of the original plan. A commenter pointed out that the dot-matrix style looks worse until you determine whether it is planning ahead or revising as it writes. This distinction is crucial for robotics tasks and equally so for agent evaluation. A high number of tool calls could reflect fine-grained control, or it could mean no continuous plan has been formed.

There is also a counterintuitive point: coding ability rankings do not necessarily transfer to motion planning. Being strong at writing code does not mean a model can turn “writing letters” into a clean robotic-arm trajectory. Benchmarks like AA-Robotics v0.1 pull models from text interfaces into physical action sequences and expose a different kind of capability. They do not replace coding or reasoning benchmarks, but they add a very concrete area of observation: whether a model can break a goal into reasonable actions, reduce unnecessary round trips, and maintain a plan across a continuous task.

What is currently public is AA-Robotics v0.1, and more examples are in the original link. It is not suitable for directly giving the two models an overall ranking. But it provides a very clear demonstration: for the same “hello world,” one model treats it as 3,270 points, and another treats it as 514 strokes. The number of actions, the time taken, and the handwriting are all in the video, so viewers can judge for themselves which process is closer to “writing.”