Robotic hand grabs an apple stem: Kimi 6.8 seconds, GPT-6 Astra 13.9 seconds, 10x cost difference

A five-fingered robotic hand reaches toward an apple, pinches the stem at the top, and lifts the apple off the table. The timer stops at 6.8 seconds.

This is a set of comparison tests posted by PhysAIAnalysis. Task: use a five-fingered robotic hand to grasp the apple stem and pick up the apple. One attempt per model. The original post says both lifted it.

Here is the data side by side:

• Kimi K3.1 Pro: lifts in 6.8 seconds, 9 minutes to set up, cost 34 cents.
• GPT-6 Astra: lifts in 13.9 seconds, 11 minutes to set up, cost $3.48.

The speed difference is about 2x, and the cost difference is more than 10x. The original post also mentions that they ran 17 models on this test, with the full ranking at the end.

Grasping an apple stem is genuinely difficult. The stem is thin and its surface is slippery, so the robotic hand first has to locate it, then control grip force, and lift without letting it slip. For an ordinary person this is a casual grab; for a robotic hand it is a combined problem of perception, planning, and execution. In the video, both hands complete the action, but their pacing and posture are different.

The comments didn't wait for the ranking. Some joked about GPT's hand coordination, some asked what population GPT's hand data comes from, and someone posted a clip of "how Astra drinks coffee": in the video, a person rotates a white bowl with one hand.

The original post puts the ranking of 17 models at the end, but the comments first went to look at the hands. That reaction is not surprising. For a task like grasping an apple stem, the result is only "picked it up" and "didn't pick it up," but the process contains hand-eye coordination, tactile feedback, and cost trade-offs. 6.8 seconds and 13.9 seconds are just surface numbers; 9 minutes of setup and 34 cents are what need to go into the same table.

Original post link: https://x.com/PhysAIAnalysis/status/2116959349398778008