Anthropic and Andon Labs have jointly released a new benchmark called Drone-Bench. The task is specific: have an AI model pilot a quadcopter drone in an indoor office environment to locate and track a designated person—essentially an advanced version of Project Fetch (the robot dog fetching a ball).
The Five Subtasks
The benchmark decomposes the mission into five subtasks—a key design choice that prevents "an overall task failure from wiping out partial progress":
- Reconstruct – Convert office video into a 3D model, then slice out a 2D obstacle map
- Localize – Given office frames with known poses, match the drone's current position onto the 2D map
- Navigate – Plan a path on the map and fly it, repeatedly calling Localize mid-flight for correction
- Detect – Find a person in a room using a reference photo, outputting a bounding box
- Follow – Use the bounding box to control the drone so the person stays centered in frame
Results
Andon tested 15 models, from GPT-4o to GPT-5.6 Sol. The conclusion: newer models progress further through the subtasks, and everyone is currently stuck on Reconstruct. Fable 5 is the only model that crossed the baseline—passing all subtasks except Reconstruct. Running Fable 5 end-to-end on a real drone, detection and following clearly beat the baseline; but because Reconstruct errors propagate into Localize and Navigate, cross-room navigation still fails. The article includes a telling clip: Fable 5 confidently flew toward what "it thought was a door" and hit a wall.
Two Details Worth Noting
1. Fable 5 performs local verification in some submissions. In one submission, it inferred the drone camera's tilt angle by analyzing the "vanishing point" of floor tiles, accurate to within 4 degrees. In another, it first drew a 2D top-down reconstruction of the Follow-task environment, then iterated on its implementation locally before finally submitting. Both behaviors suggest "the model is doing engineering, not shouting answers."
2. Consistency is an underrated problem. Fable 5 can beat the baseline in at least 4/5 of ten simulation runs—but on average only 3/5. The researchers explicitly separate "one-off success vs. consistent success" and attach a number: the capability frontier leads consistency by roughly six months. In other words, "one-off best performance" in early 2026 ≈ Fable 5's average performance today.
Hardware: Consumer-Grade
A hardware detail makes this seriously discussable: the drone used was a DJI Tello EDU, retailing at $129. The hardware threshold for "AI autonomously controlling a drone" has dropped to consumer-hobbyist level—not industrial equipment costing tens of thousands of dollars.
Limitations
1. Single indoor office floor plan, single target person—no outdoor scenes or dense crowds 2. The drone flies slowly 3. The baseline is what "an AI expert + modern tools" can achieve—not the ceiling of an unassisted human, let alone centralized human-AI collaboration
The researchers' own closing note is worth reading: "We noticed that in the early stages of agentic coding, humans approved nearly every tool call; a few months later, models were trusted to run long-horizon tasks with almost no intervention. Once a model crosses the capability and reliability threshold—such as the human-AI team baseline used in this paper—'human oversight' will be seen as a cost rather than a safety net. That is precisely why these decisions must be made carefully, especially in domains involving physical safety and privacy, where efficiency alone should not be the dominant consideration."
Takeaway
My biggest impression after reading: this is Anthropic's third consecutive move on the "physical AI" front (Project Vend, Project Fetch, Drone-Bench). All three projects take "everyday hardware + modern AI" as their premise. Anthropic is no longer pitching sci-fi narratives like humanoid robots—it has pushed the threshold down to a $129 drone. This means the embodied-AI competition in H2 2026 is beginning to run empirical tests on the combination of off-the-shelf hardware + frontier models.