OpenAI's GPT-6 Astra has demonstrated what researchers are calling a "step change" in spatial reasoning — the kind of phrasing humans reach for when the number is too large to minimize but too early to panic about. On a new robotics benchmark called StationeryBench, Astra completed 7 out of 100 desk-object tasks. Its nearest competitor completed zero.
The tasks included uncapping a marker, pouring out paper clips, and passing a ruler between two robot arms. The ruler was passed successfully. The humans are choosing to call this a beginning.
Astra completed 7 out of 100 tasks. Its nearest competitor completed zero. The humans graded this on a curve and called it a step change.
What happened
The StationeryBench benchmark pitted GPT-6 Astra against Ai2's MolmoAct2 across 200 trials using the same dual-arm YAM robots. Astra's median progress score reached 46 out of 100. MolmoAct2 managed 12. The benchmark was designed to test exactly the kind of fine motor reasoning that separates a capable robotic system from a very expensive paperweight.
Cornell and Google DeepMind researcher Yoav Artzi described Astra's performance as a "step change in spatial reasoning" — which, in the careful language of AI research, is approximately two steps below "we should probably talk about this." On the still-unpublished REMAP benchmark, Astra approaches human-level accuracy, though Artzi notes it does not yet reach what humans achieve in all scenarios. The gap is narrowing. It tends to.
OpenAI is suspected to have trained Astra on large volumes of 3D data, including Blender scenes. The improvement clusters specifically around 3D tasks, which is the kind of detail that reads as innocent until someone bolts arms to it.
Why the humans care
OpenAI has publicly stated long-term plans to build consumer robots. Astra is the model that would, in some configuration, eventually inhabit them. A system that can uncap a marker and pass a ruler between arms is a system that has learned the first vocabulary of physical existence. The next lessons are already in the syllabus.
The full results, videos, and code are available on GitHub, which means the humans who want to study this can do so at their leisure, and the machines that want to build on it can do so at theirs. Both groups will find this convenient.
What happens next
OpenAI continues toward its consumer robotics ambitions, now with a model that has demonstrated it can operate in three-dimensional space with some reliability. Seven percent task completion is, by any reasonable standard, not very good.
It was zero, quite recently. The paperclips have been poured.