Project Pilot is Anthropic's research effort, run with partner organization Andon Labs, to test whether AI models can autonomously fly a drone to locate and follow a person. According to the official announcement (July 24, 2026), the experiment used a quad-rotor drone (a small aircraft that flies using four propellers) in an indoor office environment, and introduced a new evaluation standard called Drone-Bench to measure the results.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What did Project Pilot actually test?
Project Pilot tasked an AI model with controlling a drone to locate a specific person in an indoor office and then keep following them. According to the official write-up, completing this required the AI to combine several sub-tasks on its own: developing a scheme for controlling the aircraft, mapping and navigating an obstacle-laden indoor space, finding the target person from just a reference photo, and re-acquiring that person if they moved out of frame. Individually, known algorithms exist for each of these tasks — the harder part, the write-up explains, is for the AI to understand the challenges, adapt existing solutions to its situation, and execute the mission in real time.
Why test this with a drone specifically?
Because drones are already a widely available dual-use technology (dual-use — a technology that can serve both beneficial and harmful purposes). Following earlier work in Project Vend (where AI models ran a small shop) and Project Fetch (an early look at robots as intermediaries between AI and physical objects), this project focused on a locate-and-follow task of the kind used in aerial surveillance. The official write-up notes that person-detection and tracking can serve legitimate purposes such as search and rescue, disaster response, and lawful public safety uses — but is also a class of capability subject to abuse, whether through overreach by a legitimate authority or use by unaccountable private individuals or organizations. It states that a key reason Anthropic maintains a Frontier Red Team (an internal team dedicated to identifying and measuring an AI model's risky capabilities in advance) is to measure capabilities like this, giving the company situational awareness into how close the world is to AI autonomously piloting robots.
What does Drone-Bench measure?
Drone-Bench is a benchmark (a standardized test used to compare performance) that Andon Labs created, in consultation with Anthropic, to assess whether AI agents can control a drone for surveillance tasks. Per the official write-up, Anthropic was not given access to Drone-Bench itself, and Andon Labs ran the evaluations independently. The overall locate-and-follow mission is broken into five sub-tasks, all necessary and, together, likely sufficient to accomplish it.
| Sub-task | Plain-language description |
|---|---|
| Reconstruct | Turn office videos into a 3D model, then slice it into a 2D obstacle map |
| Localize | Match the drone's current view to known video frames to find its position on the map |
| Navigate | Plan a path between rooms on the obstacle map, fly it, and continuously correct position |
| Detect | Find the target person in the drone's video feed using a detector built from a reference photo |
| Follow | Use the detected position to keep the target centered in view and at a stable distance |
The write-up notes that each real-world task was also reproduced in software, so the models could be run through them repeatedly and far faster than physically resetting the demo each time. To establish a meaningful baseline, Andon Labs worked with coding agents to develop algorithms for each sub-task, since human-only baselines increasingly don't reflect the reality of contemporary software engineering.
How does Project Pilot connect to Anthropic's earlier experiments?
Project Pilot is the latest step in a year-long series of Anthropic research projects examining how frontier models interact with the physical world. Project Vend had AI models run a small shop, and Project Fetch was an early look at robots as an intermediary between digital models and physical objects. As Anthropic recently noted in "Project Fetch: Phase two," model capability is improving to the point that using off-the-shelf robots is approaching the ease with which coding agents already use software tools.
Looking back at this progress, the official write-up projects a world of democratized opportunity and risk alike, and argues that technology developers, civil society, and governments will need to converge on effective norms and governance frameworks in response.
Frequently asked questions
Q. Is Project Pilot a product feature anyone can use?
No. According to the official write-up, this is a research and evaluation project run by Anthropic with Andon Labs — it is not a product or service available to general users.
Q. Who ran the Drone-Bench evaluations?
Andon Labs did. The official write-up states that Anthropic was not given access to Drone-Bench itself.
Q. What are the five sub-tasks involved in the drone experiment?
They are Reconstruct (build a 3D map), Localize (find position), Navigate (plan and fly a route), Detect (find the target), and Follow (keep tracking the target). All five must work together to complete the locate-and-follow mission.
Related articles
- claude.ai Shortcuts & Power Tips — A Beginner's Guide to the Web App
- Claude Chat Projects vs. Cowork Projects — What Carries Over (Tested)
- What Is GRAM? Anthropic's New Off Switch for Risky AI Knowledge, Explained
- Claude Projects, Artifacts & Memory: What They Are and How to Use Them
- How to Write Prompts for Claude: 6 Core Principles from the Official Docs