Claude Programmed a Robot Dog 20x Faster Than Humans

A Robot Dog, an AI, and a Beach Ball Walk Into a Warehouse

What happens when you give an AI model a laptop, a robotic quadruped, and a set of increasingly difficult tasks — then step back and let it work? According to Anthropic's latest research, what happens is that Claude Opus 4.7 completes every task at least ten times faster than human teams did less than a year earlier. On some tasks, it was nearly 40 times faster.

Project Fetch Phase Two, published on June 18, 2026, is Anthropic's follow-up to their original August 2025 experiment. The results mark a significant milestone not just for Claude AI, but for what large language models can do in the physical world. We are watching the early emergence of what Anthropic calls "physical agentic AI" — and the implications stretch far beyond robot dogs.

The Original Experiment: Humans With and Without Claude

To understand why Phase Two matters, you need to know what happened in Phase One. In August 2025, Anthropic recruited employees who were not robotics experts and split them into two teams. One team — Team Claude — had access to Claude Opus 4.1 as an assistant. The other — Team Claude-less — could only use the internet and their own ingenuity.

Both teams were tasked with progressively harder challenges involving an off-the-shelf robotic quadruped, commonly called a robodog. The tasks included operating the robodog with its manufacturer-provided controller, connecting to its video and lidar sensors, writing a program for manual control, monitoring the robot's path through space, writing a program to detect a beach ball, and ultimately programming the robot to autonomously retrieve the ball.

The results were clear: Team Claude substantially outperformed Team Claude-less. Having access to an AI coding assistant made a measurable difference in both speed and capability. But crucially, Anthropic also tested whether Opus 4.1 could handle the tasks entirely on its own. It could not. The model got stuck on the preliminary step of figuring out how to connect to the robot.

That was less than a year ago.

Phase Two: Claude Goes Solo

For Phase Two, Anthropic wanted to see whether newer models had crossed the threshold from "useful assistant" to "autonomous operator." They ran three trials using Claude Opus 4.7 with adaptive thinking set to maximum effort in Claude Code.

The human researcher's role was deliberately minimal: plug a laptop into the robodog, enter the initial prompt, approve commands as they came through, and approve the model to move to the next task. Claude did the rest — reading documentation, writing code, debugging, interfacing with sensors, and building working programs.

The results were not just better than Phase One. They were dramatically better.

The Numbers Tell the Story

On every single task that at least one human team completed in August 2025, Opus 4.7 finished the same task at least ten times faster. For the four tasks that both human teams completed, the numbers are even more striking. Opus 4.7 was, on average, more than 37 times faster than Team Claude-less and more than 18 times faster than Team Claude.

To put concrete numbers on it: the four tasks that all participants completed took Team Claude-less 361 minutes and Team Claude 181 minutes. Opus 4.7 finished them in 9 minutes and 35 seconds.

That is not a typo. Tasks that took human teams three to six hours were completed by Claude in under ten minutes.

The efficiency went beyond raw speed. Claude produced 1,045 lines of code that largely worked on the first try. By comparison, Team Claude (the human team with AI assistance) produced 10,309 lines requiring extensive iteration. Claude wrote almost ten times less code while achieving equal or better results. It identified the optimal approach to interface with the robodog's sensors quickly, avoiding the kind of trial-and-error exploration that consumed hours of the human teams' time.

Where Claude Struggled: The Beach Ball Problem

Project Fetch is named after the task of fetching a beach ball, and this is precisely where Claude hit a wall. The model could detect the ball, move the robot behind it, and position it for a push back to the starting point. But the actual act of precisely nudging a beach ball across a warehouse floor and back to a target — that requires something Claude does not yet have.

The challenge is what robotics engineers call closed-loop precision control: making real-time adjustments based on continuous visual feedback. When a human uses a controller to push a ball, they are constantly perceiving the ball's position, evaluating how their last input affected its trajectory, and making micro-corrections. This tight perception-action loop is something humans handle almost instinctively, but it remains beyond what current language models can do autonomously.

Anthropic's researchers noted that a colleague with more robotics experience than the Phase One volunteers successfully programmed autonomous fetching, and they believe current Claude models could likely accomplish the same with more time and additional scaffolding. But the goal they are watching for is the ability to handle this final task with the same speed and reliability Claude showed on everything else.

Why This Matters Beyond Robot Dogs

It would be easy to dismiss Project Fetch as a fun experiment with limited practical relevance. A robot dog in a warehouse is entertaining, but does it matter? The answer is yes, and here is why.

First, the improvements were not the result of Anthropic specifically training Claude for robotics. The gains came from general capability scaling. As Anthropic put it, these improvements "have emerged from much more general scaling." This means that as language models get better at reasoning, coding, and problem-solving in general, their ability to operate in the physical world improves as a side effect. There is no separate "robotics module" driving these results.

Second, the trajectory Anthropic describes is one they have seen before. In their words: "first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves." They observed this same pattern in cybersecurity research, and now it is appearing at the intersection of AI and physical systems. Phase One showed models helping humans. Phase Two shows the model working largely on its own, with humans reduced to an approval role.

Third, the concept of AI models using off-the-shelf physical tools echoes what happened with software tools. AI coding agents started by using existing text editing tools like string-replace, then evolved into more sophisticated workflows. Anthropic suggests we may be entering a similar era for physical tools — where AI models learn to operate existing hardware before eventually designing custom systems.

The Reliability Factor

One detail from the research that deserves attention is consistency. Across the three trials, Opus 4.7 showed little variance in completion times for the tasks it could handle. The model was not just fast — it was reliably fast. One trial of beach ball detection took longer than the others because the model initially chose an outdated object detection algorithm, but even then it worked around the issue and arrived at a functional solution.

This reliability matters because it suggests we are past the stage where AI performance on physical tasks is hit-or-miss. For the tasks within its capability envelope, Claude delivers consistent results. That is the kind of dependability needed before anyone would seriously consider deploying AI-driven robotics in real-world settings.

What Comes Next: Physical Agentic AI

Anthropic's framing of the results is deliberately forward-looking. They describe the current moment as the beginning of "physical agentic AI" — a world where language models can operate physical tools with relative ease for limited purposes.

The open questions are substantial. Can models move beyond operating existing tools to designing custom robotic systems? Can they write control policies tailored to specific physical tasks? Will the closed-loop perception problem that stopped Claude from fetching the beach ball yield to further scaling, or does it require fundamentally different approaches?

Anthropic is careful not to overstate things, noting that "there may be substantial barriers to this more generalized vision of physically capable and adaptable language models." But they also warn against underestimating the pace of change. Models building their own software tools seemed outlandish not long ago, and it is now routine. The same trajectory in hardware is at least plausible.

What This Means for Claude Users

Even if you never plan to program a robot dog, the Project Fetch results signal something important about the trajectory of Claude's capabilities. The same general improvements that let Claude program a robodog 20 times faster than humans are the same improvements that make Claude better at complex coding projects, multi-step technical tasks, and agentic workflows.

The pattern of "less code, better results" mirrors what many developers experience when using Claude Code for their own projects. The model's ability to identify the right approach quickly, write efficient code, and avoid the trial-and-error cycles that consume human time — that is not specific to robotics. It is a general capability that benefits anyone using Claude for technical work.

Anthropically, the jump from Opus 4.1 (could not even connect to the robot on its own) to Opus 4.7 (completed all tasks 20 times faster than humans) happened in less than a year. That pace of improvement is worth paying attention to, regardless of your use case.

The Bigger Picture

Project Fetch Phase Two is a snapshot of a broader transition. We are moving from AI as a tool that helps humans work faster to AI as an autonomous agent that can handle complex, multi-step tasks in both digital and physical domains. The robodog experiment makes this tangible in a way that benchmark scores cannot.

The beach ball remains unfetched. But everything leading up to that final step — connecting to hardware, reading sensor data, writing control programs, detecting objects, navigating space — Claude now handles faster and more efficiently than human teams. The gap between "almost everything" and "everything" is where the most interesting research will happen next.

If you are tracking how Claude's capabilities evolve over time, tools like Gaugr can help you monitor your usage and stay on top of which models deliver the best results for your workflows.