Duan runs Astra, Opus, Grok and MolmoAct2 head to head, unedited
Four models, one task, real time, and Chris Paxton is not impressed with how long the Astra rollout takes.
Jiafei Duan posted a four-way comparison of GPT-6 Astra, Opus 5.5, Grok 4.7 and MolmoAct2 running the same manipulation task. His framing is the point. Real time, no sped-up demos, zero-shot (no task-specific training first), no cherry-picking.
Real time. No sped-up demos. Zero-shot. No cherry-picking.
That follows his earlier clip of a single GPT-6 Astra rollout controlling the MolmoAct2 YAM arms, which he posted to show how slow the thing actually is when nobody trims the footage. He said then that there is potential but he does not think it is practical for real-world robotics deployment yet.
What Duan takes from it
His read on the four-way run is that LLM robot control is progressing but is not there yet, and that low-level control still matters, which he ties to Moravec's paradox, the old observation that the things humans find easy are the hardest for machines.
He also flagged one detail that is easy to miss. When MolmoAct2 moved its left arm to place the candy, the right arm moved away to give it space. Nobody asked it to do that.
Paxton's verdict
Chris Paxton was blunt about the Astra rollout. He called it incredibly slow for a very simple pick and place task, and said you could probably vibe code it faster with SAM. He immediately qualified that, saying SAM is not a real solution either because it is not reliable enough for real environments and tasks. His conclusion is about the field, not the model.
Incredibly slow for a very simple pick and place manipulation task that you could probably vibe code faster with SAM.
"But robotics benchmarking largely remains not challenging enough!"
That is the argument doing the rounds. If a general-purpose language model can be wired up to arms and score on a benchmark while taking minutes to move one object, the benchmark is measuring something too easy.
Shown versus claimed
What is actually shown here is four rollouts at true speed on one task. Duan is not publishing success rates across a suite, and the sources do not give scores or run counts for the comparison. What makes it useful is the thing missing from most footage of these systems, which is the waiting. A sped-up clip of any of these four would look competent. Played at the speed it happened, the gap between a model that can reason about a scene and a model that can act in it is obvious without anyone having to argue about it.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.