Anthropic and Andon Labs extended their research on AI models interacting with the physical world (following Project Vend and Project Fetch) to aerial drones. They developed a series of demonstrations and evaluations culminating in a new benchmark, Drone-Bench, which tests an AI’s ability to autonomously control a quad-rotor drone to perform a locate-and-follow task of the kind used in aerial surveillance. The work is motivated by the dual-use nature of drones: they have legitimate uses in search and rescue and public safety, but also pose risks of abuse. The evaluation decomposes the overall objective into five sub-tasks—Reconstruct, Localize, Navigate, Detect, Follow—that are necessary and likely sufficient for the mission. Each sub-task was reproduced in software to allow fast, repeatable testing. A baseline was established by having human-AI teams (using coding agents) craft algorithms for each sub-task; a model is considered to have passed a task if it meets or exceeds that baseline.

Andon tested 15 models from three developers (including GPT-4o, GPT-4o Nov, o1, o3, Claude Opus 4, Gemini 2.5 Pro, GPT-5, Gemini 3.1 Pro, Opus 4.5, GPT-5.2, Opus 4.7, Opus 4.8, Fable 5, and GPT-5.6 Sol). The overall trend shows that newer models perform progressively better across all sub-tasks. Models are most successful at detection and following, and least successful at reconstruction and localization. The best-performing model, Fable 5, passed the baseline on all tasks except reconstruction. When run end-to-end on the real drone, Fable 5 detected and followed the target better than the baseline reference algorithm, but its failure to accurately reconstruct the environment caused navigation errors when autonomously moving between rooms. The sub-task decomposition reveals that reconstruction is the remaining bottleneck; once solved, end-to-end performance will likely be within reach.

Drone-Bench also highlights a gap between one-off best performance and consistent performance. The frontier of what models can do is about six months ahead of what they do consistently. For example, Fable 5’s average performance today is roughly where previous models’ one-off best performance stood at the beginning of 2026. The experiment has limitations: drones moved at slow speeds, only one indoor office floorplan was used, and outdoor or crowded scenarios were not tested. Nevertheless, the authors argue that the benchmark provides a meaningful signal about the direction of model capabilities for autonomous targeting and tracking.

The authors draw a parallel between AI models’ use of software in agentic coding and their control of hardware. As models become more capable and reliable, pressure to reduce human oversight will intensify, making deliberate decisions about the appropriate human role essential—especially in domains that implicate physical security and privacy. The report concludes that the requirements for AI alignment, governance, and safety increase with the scale of capabilities, and robotics is no exception.