An AI model has outpaced human teams on physical-world robotics programming tasks by a factor of 20 — and according to Anthropic, the advance arrived not through targeted robotics training, but as a byproduct of the same general capability scaling that has driven large language model improvements across every other domain.
The finding comes from Project Fetch Phase Two, the second installment of an experiment Anthropic’s Frontier Red Team first ran in August 2025. In that original test, two groups of non-expert Anthropic employees competed to program an off-the-shelf robotic quadruped — a so-called “robodog” — using sensor-interfacing and computer vision tasks. One team used Claude Opus 4.1; the other relied only on the internet and their own problem-solving. The AI-assisted team finished more tasks faster. When Anthropic checked whether Opus 4.1 could complete the tasks entirely on its own, it failed at the very first step — connecting to the robot.
Ten months later, Claude Opus 4.7 not only connected but finished everything that both human teams completed, and it did so in 9 minutes and 35 seconds. The Claude-assisted human team from August 2025 needed 181 minutes. The team working without AI needed 361 minutes. That gap — 19 times faster than the AI-assisted humans and nearly 38 times faster than the unassisted group — makes this the most precisely quantified speed comparison yet between an autonomous large language model agent and trained human teams performing real-world physical tasks.
What Opus 4.7 Did and How Its Agentic Architecture Made It Possible
The four tasks the model tackled matched the subset completed by both human teams in the August 2025 experiment: connecting to the robodog’s video camera sensor, connecting to its lidar sensor, writing a program to monitor the robot’s path through space, and detecting a beach ball using computer vision. A researcher’s role was limited to plugging a laptop into the robot, entering the initial prompt, approving commands, and signaling the model to advance to the next task.
The architectural mechanism behind the performance advantage is specific. Claude Code runs on what Anthropic calls the agentic loop — a three-phase cycle in which the model gathers context, takes action, and verifies results, then repeats. Tools are what make the loop agentic: without them, the model can only respond with text; with them, it can read sensor outputs, write and execute code, observe whether a sensor connection succeeded, and correct its next command accordingly. For Project Fetch Phase Two, Opus 4.7 ran with adaptive thinking at maximum effort — a reasoning mode that lets the model think between individual tool calls, not just before them. That interleaved reasoning is what enables reliable performance across multi-step physical tasks: the model sees an intermediate result, adjusts its next action, and continues without halting for human input.
The code efficiency figures are as telling as the speed figures. Opus 4.7 produced 1,045 lines of code across all tasks, compared to the 10,309 lines the Claude-assisted human team wrote in August 2025. The human team’s higher volume reflects iterative, exploratory coding: trying one sensor API, failing, trying another, accumulating dead ends. Opus 4.7 identified the most direct integration path for each task on the first attempt in most cases and wrote only the code the task required. Much of what it produced worked immediately.
The model ran three trials per task. Performance times across those runs were highly consistent — a signal of reliability rather than luck. One beach ball detection trial ran substantially longer because the model initially selected an outdated object detection algorithm, but it recognized the problem, worked around it without human intervention, and arrived at a correct solution.
The Architectural Wall: Where Closed-Loop Control Stops LLMs Cold
The one task the model did not solve is the most instructive finding in the experiment. The “fetching” portion — physically guiding the robodog to nudge a beach ball back to its starting position — requires real-time closed-loop control. That means the system must continuously read sensor data about where the ball went after each push, compute how the robot’s last command produced that result, and issue corrective commands faster than the situation changes.
This is a structurally different category of problem from what Opus 4.7 succeeded at. Programming sensor connections, writing path-monitoring code, and implementing computer vision are all open-loop tasks: the model writes instructions, executes them, and observes a result. Closed-loop control requires the system to be in continuous communication with the environment — receiving sensor feedback and acting on it in near-real time, at a speed and latency that current large language model inference architectures do not support.
Human participants, after practicing with a handheld controller, could do this naturally. They perceived the ball’s drift, interpreted the error from their previous command, and corrected. Opus 4.7 could position the robot correctly behind the ball but could not execute the rapid, adaptive adjustments the task demands.
Anthropic notes that one of its researchers with more robotics experience than the Phase 1 volunteers did successfully write a program for autonomous ball retrieval, establishing that the closed-loop problem is not structurally unsolvable — only beyond what the current generation of models achieves through general scaling alone.
Claude AI Robotics Gains Come From General Scaling, Not Robotics Engineering
One of the most consequential claims in the Phase Two paper is also the easiest to pass over. Anthropic states explicitly that the gains observed — from failing at sensor connection in August 2025 to finishing all four tasks in under ten minutes by June 2026 — did not result from any deliberate effort to improve Claude’s robotics capabilities. Anthropic’s researchers write that “these improvements, like so many others in the history of LLM development, have emerged from much more general scaling.”
That claim carries a specific prediction. When AI models first gained the ability to handle software coding tasks autonomously, they did so by learning to use existing software tools — text editors, bash terminals, file readers — the same tools human developers already used. The hypothesis the paper develops is that a comparable transition is now beginning for physical tools: off-the-shelf robotic hardware is becoming one more category of tool that a capable LLM can learn to use through general capability improvement, not targeted robotics engineering.
Anthropic calls this the beginning of “physical agentic AI.” If general scaling moved Opus 4.1 — unable to connect to a robot — to Opus 4.7 — 19 times faster than human teams on four tasks — in under a year, then the question of when closed-loop physical control becomes achievable is, under this framing, a question about when general scaling produces the required capability, not about when someone builds a purpose-built robotics model.
The same three-stage progression Anthropic previously documented in cybersecurity — models first augment humans, humans briefly guide models, models then operate independently — appears to be taking shape at the physical layer of robotics.
What Claude Code Physical AI Autonomy Means for the Broader Field
Experts at the World Economic Forum’s January 2026 annual meeting in Davos reached a specific consensus: the foundational era of robotics is over and the deployment era has begun. Daniela Rus, director of the Computer Science and Artificial Intelligence Laboratory at MIT, confirmed at the session that entire fleets of robots now operate around the clock moving shipping containers without human intervention. Those systems rely on purpose-built robotics software — not on a general-purpose large language model using an agentic coding tool.
What Project Fetch Phase Two adds to that picture is a data point about rate of change. The experiment’s design is deliberately constrained: an off-the-shelf robodog, non-expert researchers, a commodity laptop, the standard Claude Code interface, no robotics-specific engineering. The result was a model that outperformed even AI-assisted human teams at the programming tasks that constitute the first and most demanding phase of robot deployment — getting the device connected, sensed, monitored, and vision-capable.
The Anthropic Economic Index report published June 26, 2026, frames this as part of a broader pattern of AI capability maturation visible across multiple domains simultaneously. The robotics finding appears alongside data showing Claude now writes more than 80% of the code merged into Anthropic’s own codebase, and that engineers at Anthropic are merging eight times as much code per day as they were in 2024. The Physical AI transition, per Anthropic’s framing, is one more instance of the same underlying dynamic playing out at the physical layer.
Frequently Asked Questions
How much faster is Claude AI at robotics programming than human teams?
In Anthropic’s Project Fetch Phase Two experiment, Claude Opus 4.7 completed four robotics programming tasks in 9 minutes and 35 seconds. The Claude-assisted human team from August 2025 needed 181 minutes for the same tasks; the unassisted team needed 361 minutes. That works out to roughly 19 times faster than the AI-assisted team and nearly 38 times faster than the team working without AI. Across all tasks completed by at least one human team, the model was at least 10 times faster in every case.
What can Claude AI still not do in agentic AI robotics?
Closed-loop real-time control remains beyond the current generation of large language models. Tasks such as nudging a ball to a specific location require the system to continuously read sensor feedback about each action’s outcome and issue corrective commands faster than the situation changes — a feedback loop that runs at a speed current LLM inference latency cannot match. Open-loop tasks (writing programs to connect sensors, monitor paths, and detect objects) are where LLMs now excel; real-time closed-loop physical control is where the architectural gap persists.
What is physical agentic AI and why does Project Fetch matter for LLM robotics programming?
Physical agentic AI refers to AI systems capable of autonomously programming and controlling physical hardware — robots, actuators, and sensors — using the same kind of tool-use architecture that agentic AI already applies to software tasks. Project Fetch Phase Two is the first documented case in which a general-purpose large language model, operating through Claude Code with minimal human oversight, outperformed trained human teams on physical-world robotics programming tasks at this margin. Anthropic argues the same general capability scaling that drove software AI gains is now beginning to extend into physical tool use.
What is the open-loop versus closed-loop distinction in AI robotics, and why does it define where LLMs succeed and fail?
Open-loop tasks require the AI to write and execute a program and observe whether it worked — the model is not in real-time communication with the physical environment between commands. Closed-loop tasks require the model to continuously read sensor data, compare it to the desired state, and issue corrections faster than the environment changes. Large language models now perform open-loop robotics programming at speeds that exceed human teams. Closed-loop real-time control requires an inference speed and continuous-feedback architecture that current LLMs do not yet provide.