GPT-6 Astra Tops New Robotics Benchmark as Rival Model Scores Zero

OpenAI's GPT-6 Astra has posted the first measurable gains in spatial reasoning for a frontier language model, according to results from a new robotics benchmark called StationeryBench. The model completed 7 of 100 tasks using dual-arm robot hardware, a score that looks modest in isolation but stands out because the competing model, MolmoAct2, failed to finish a single task. A researcher involved with the evaluation described the gap as a step change in spatial reasoning, a phrase that will draw scrutiny from a field long accustomed to benchmark claims outrunning real capability.

StationeryBench is designed around manipulation chores that require a model to locate objects, plan a grip, and coordinate two arms in physical space. That combination has been a persistent weak point for large models, which can describe a scene in fluent prose yet misjudge which side of a cup the handle sits on. Seven successful runs out of 100 is not autonomy by any normal definition, but the zero score from MolmoAct2 means the comparison is not a rounding error. The result suggests Astra is converting language-level world knowledge into motor commands at a rate previous systems did not reach.

The detail that matters most is the hardware. Dual-arm setups introduce coordination problems that single-arm benchmarks hide, because the model has to decide which limb acts, in what order, and how to avoid the two arms colliding. Most published robotics evaluations sidestep this by testing grasping in isolation or by using scripted motion primitives. StationeryBench appears to force the model to handle the sequencing itself, which is why a 7 percent success rate is being treated as newsworthy rather than trivial.

Independent verification is the obvious missing piece. The results come from an early benchmark run, and the summary offers no breakdown of how many attempts each task allowed, how the 100 tasks were selected, or whether the model was fine-tuned on similar environments. Robotics scores are notoriously sensitive to those choices, and a single failed run can swing a small sample. Until the evaluation protocol and the raw logs are public, the step-change framing rests on one comparison against one competitor.

What to watch next is whether the gains survive harder tests. If Astra's spatial advantage holds on longer task horizons, cluttered scenes, and hardware it was not tuned for, the result points to a real shift in how language models ground physical action. If the 7-out-of-100 figure turns out to be a benchmark artifact, it joins a long list of robotics headlines that did not replicate. Either way, StationeryBench has given the field a number worth arguing about.