OpenAI focused much of the GPT-6 Astra release on computer use. Models operating computers need to locate interface elements, read visible text, and understand what each element does. Those requirements also improve the skills needed for computer vision.To measure those capabilities, we ran Astra through the Roboflow Vision Evals. The benchmark covers common vision tasks, including object detection, counting, and visual reasoning. We also tested capabilities outside the benchmark, including box ...