OpenAI’s GPT-6 Astra reportedly achieves 95% success rate in robotic arm control

leaderobot.com reported on Sept. 17 that OpenAI announced GPT-6 Astra in early September. According to the article, the model can use robotic arms to directly control actions such as picking up and placing objects, without additional training. In real-world tests by independent evaluation organization Robocurve, a dual-arm robotic system achieved a 95% success rate on a task involving picking up a cube and putting it in a bowl, reportedly outperforming Claude Fable 5.1, which scored 40%.

If the information reported by leaderobot.com is accurate, evaluating robot control requires looking beyond the performance of the hardware itself and considering both the generalization capabilities of foundation models and real-world testing. As large-model companies such as OpenAI move into direct robotic arm control, the key question is whether the main battleground in embodied AI will shift to the model layer.

What stands out in this report is that the benchmarks for embodied AI are shifting beyond the sophistication of the robot itself toward the generalization capabilities of foundation models. According to the report, GPT-6 Astra can operate robotic arms without additional training and achieved a high success rate on a relatively routine pick-and-place task. The new development is the growing focus on the idea that large models could connect directly to the physical world, rather than discussing robot control only within dedicated software stacks. Our related article, Unitree unveils H2 humanoid robot measuring 180 centimeters tall and weighing 70 kilograms, described advances on the hardware side, focusing on size and smoothness of movement. This article, by contrast, presents a picture in which the brain-like layer outside the hardware performed more strongly in real-world benchmarks. In other words, the competitive axis in the robotics industry could shift from hardware performance to which model controls the higher-level functions. However, the tasks reported so far are mainly relatively limited ones, such as picking up and placing objects. Whether the same performance holds for complex contact manipulation or stable, long-duration control is a separate question. The effects would extend beyond robot manufacturers to the broader industry structure. The validity of evaluation organizations such as Robocurve, data collection and annotation supply chains, and model designs that divide responsibility for spatial understanding and physical control will all come under scrutiny. In China, reactions include both voices acknowledging the advantage of leading foundation-model companies and others pointing to the limits of physical generalization. For now, it remains unclear whether large models alone are sufficient or whether a dedicated control layer is still needed. The next things to check are reproducibility on tasks outside public benchmarks, performance across different hardware and camera conditions, and the scope of the training data.

For Japanese robotics companies, competitiveness could depend not only on hardware design and control systems but also on integration with higher-level models that incorporate language and spatial understanding. However, the report confirms success rates only on limited tasks; whether the same results can be achieved with Japanese systems requires separate verification.