OpenAI’s upcoming GPT-6 Astra model is demonstrating a significant leap in physical world understanding, specifically in spatial reasoning. In early tests using the StationeryBench robotics benchmark, Astra controlled dual-arm YAM robots to perform delicate desk tasks like uncapping markers and passing rulers. The results were stark: across 200 trials, Astra fully completed 7 out of 100 tasks. Its competitor, Ai2’s MolmoAct2, completed zero. Astra achieved a median progress score of 46 out of 100, while MolmoAct2 lagged far behind at just 12. These numbers suggest Astra isn’t just slightly better; it is operating on a different plane of capability when interacting with 3D space.

Yoav Artzi, an AI researcher at Cornell and Google DeepMind, describes this performance as a “step change in spatial reasoning.” His assessment carries weight given his expertise in multimodal AI. On the unpublished REMAP benchmark, Astra’s accuracy approaches human levels, though Artzi cautions that it still falls short of human proficiency in more complex scenarios. He suspects OpenAI trained Astra on massive datasets of 3D environments, such as Blender scenes. This hypothesis aligns with the model’s specific strengths in navigating three-dimensional tasks, indicating a strategic shift in how foundation models are being prepared for physical interaction.

This development signals that the race for embodied AI is accelerating faster than many anticipated. While text and image generation have dominated headlines for years, the ability to manipulate physical objects requires a fundamentally different kind of intelligence. Astra’s performance suggests OpenAI is successfully bridging the gap between digital knowledge and physical action. The fact that all results, videos, and code are available on GitHub adds a layer of transparency that allows the broader research community to verify these claims independently.

OpenAI has long-term ambitions to build consumer robots, and benchmarks like StationeryBench provide a glimpse into that future. If Astra can master simple desk tasks now, more complex household chores may be within reach sooner than expected. However, the gap between uncapping a marker and folding laundry remains vast. The industry will be watching closely to see if this “step change” translates to real-world reliability or if it remains a controlled laboratory achievement.