Anthropic Tests 15 Models Flying a Surveillance Drone

Anthropic Put 15 Models in Charge of a Surveillance Drone

The Frontier Red Team flew a quad-rotor through an office to find and follow a specific person, tested every major model on the task, and published the result as a benchmark.

Jul 25, 2026Source: Anthropic

Anthropic's Frontier Red Team, working with Andon Labs, handed 15 AI models the controls of a real quad-rotor drone and asked them to fly through an office, find a specific person and follow them. The results and the task suite were published on Friday as Drone-Bench.

How it was scored

The task was decomposed into five sub-tasks, including Reconstruct — turning office video into a 3D model — and Localize, matching the drone's view to a 2D obstacle map. The comparison point is not an unaided human: the baseline was set by human-AI teams using coding agents. Any "models match humans" framing misdescribes what was measured.

Where the wall is

Detection and following reached near 100% of baseline. Reconstruction reached only about 47% by mid-2026, and it is the bottleneck holding the whole task back. The strongest model beat the baseline on every sub-task except reconstruction — but even it cleared the baseline on average for only three of five.

Stated limits

Anthropic published the constraints rather than burying them: slow drone speeds, a single office floorplan, a limited number of people, no outdoor testing and no crowds. Models tested spanned 15 releases from three developers across roughly two years of progress.

Which models

The 15 span roughly two years of frontier progress across three developers, running from GPT-4o and o1 through Gemini 2.5 Pro and successive Claude Opus releases to the current generation. Reading them as a time series is the point: detection and following are close to saturated, while reconstruction has barely moved.

The governance line

The team named the risk directly: once reliability thresholds are passed, there will be real pressure to treat human oversight as a cost rather than a safeguard.

Why It Matters

A frontier lab publishing a dual-use physical-autonomy evaluation before the capability arrives is the opposite of how this usually goes. It also gives an unusually clean answer to the question of what is actually blocking embodied AI — not planning or control, but spatial reconstruction.