Photo: Tima Miroshnichenko / Pexels
OpenAI's latest model is the first AI to beat human performance across all five drone navigation subtasks, though the full success rate tells a more sobering story
OpenAI’s GPT-6 Astra can now fly a drone through an office, find a specific person from a photo, and follow them around. The AI handles the entire process: building a 3D map of the space, figuring out where it is, planning a route, identifying its target, and tracking that person as they move.
That’s the good news. The less glamorous part: it pulls off the complete sequence roughly 2.8% of the time.
What Astra actually did #
The demonstration was conducted by Andon Labs using a benchmark called Drone-Bench, which breaks autonomous drone navigation into five distinct subtasks: 3D mapping, self-localization, route planning, person detection, and motion tracking.
Astra became the first AI model to surpass human-level performance on all five of those individual subtasks in at least one trial run. Previous models could handle some pieces of the puzzle but couldn’t match human competence across the entire chain.
When Andon Labs ran the full end-to-end sequence across multiple attempts, the complete success rate averaged about 2.8%. To put that in perspective, if you asked Astra to perform the full find-and-follow mission 100 times, it would nail it roughly three times.
The experiment used a basic, off-the-shelf consumer drone.
Astra’s broader positioning #
GPT-6 Astra launched in early September 2026, positioned by OpenAI as a state-of-the-art model for computer use and professional workflows. The drone piloting results were reported on September 11, 2026, just days after the model became available.
The model has been shown coordinating robotic arms and handling complex game-playing scenarios, suggesting that OpenAI is deliberately pushing Astra beyond screen-based tasks and into embodied AI territory.
The five-subtask breakdown is worth appreciating for its complexity. Building a 3D map requires spatial understanding. Self-localization means the AI has to know where it is within that map, essentially solving its own GPS problem indoors. Route planning demands optimization. Person detection from a photo involves computer vision. And tracking a moving target adds real-time adaptation to the mix.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our