From being unable to connect the robot to beating it 19 times in a year: AI enters the physical world without being trained for it

By Momentum IA · June 28, 2026.

Less than a year ago, Claude Opus 4.1 failed at the first step of a robotics experiment: connecting to the robot. Ten months later, its successor, Opus 4.7, completed the same four tasks —camera connection, lidar, trajectory monitoring and detecting a ball via computer vision— in 9 minutes and 35 seconds. The AI-assisted human team in August 2025 took 181 minutes. The team without AI, 361. That is 19 times faster than the assisted humans and almost 38 times faster than those who worked alone. In terms of code efficiency, the model wrote 1,045 lines versus the 10,309 of the human team; most of it worked on the first attempt.

Those are the facts of Project Fetch Phase Two, the follow-up experiment that Anthropic's Frontier Red Team has just published. But the most important fact is not the speed: it is the cause. Anthropic states bluntly that these improvements are not the result of any deliberate effort to improve Claude's robotics capabilities. They are, literally, a byproduct of the same general scaling that has driven advances in text, code and reasoning. Robotics came along for free in the luggage.

The architecture that makes it possible deserves a moment of attention. Claude Code operates on what Anthropic calls the agentic loop: a three-phase cycle —gather context, act, verify result— that repeats continuously. What turns that loop into something capable of interacting with physical hardware is the tools: without them, the model only produces text; with them, it can read sensor outputs, write and run code, observe whether a connection worked and correct the next command. For Fetch Phase Two, Opus 4.7 ran with adaptive reasoning at maximum effort, which allows it to think between individual tool calls, not only before them. That is what enables reliable chaining in multi-step physical tasks: the model sees an intermediate result, adjusts and continues without asking for help.

Where the model failed is as instructive as where it succeeded. The 'fetch' task —physically guiding the robot dog to push a ball back to its starting position— requires real-time closed-loop control: continuously reading sensor feedback, computing the error of each previous command and issuing corrections faster than the situation changes. That is structurally different from writing a program that runs and observes. Current language models cannot withstand the inference latency that such a continuous-feedback architecture demands. Humans, after practicing with a controller, did it naturally. Opus 4.7 did not. It is a real architectural limit, not one of cognitive capability.

Our reading: the open-loop / closed-loop distinction is, at this moment, the sharpest frontier between what AI can already automate in robotics and what still needs purpose-built solutions. But that frontier is moving. And the speed of the movement is what should focus the industry's attention.

The historical pattern Anthropic documents —first the model augments the human, then the human guides the model with minimal supervision, finally the model operates independently— they had already traced in cybersecurity. Project Fetch shows that the same arc is replaying in the physical layer of robotics. And if the leap from 'cannot connect' to '19 times faster than the assisted human' happened in ten months without any robotics-specific work, the question of when closed-loop control becomes attainable is not 'when someone builds a model specifically for it', but 'when general scaling produces the required capability'. That is the hypothesis Anthropic is betting on, and so far it is right.

This fits into a broader context that Anthropic itself published on June 26 in its Economic Index: Claude already writes more than 80% of the code merged into Anthropic's own codebase, and its engineers are integrating eight times more code per day than in 2024. Physical robotics is not an isolated case; it is one more instance of the same dynamic unfolding in a new layer.

For the industry, the practical meaning is twofold. First: hardware commissioning tasks —connecting sensors, writing API integrations, implementing computer vision— that today consume weeks of specialized engineering are being reduced to minutes by a general-purpose LLM with access to standard tools. That transforms the economics of robotic deployment for companies that cannot afford dedicated robotics teams. Second: the experiment used off-the-shelf hardware, a standard laptop and the normal Claude Code interface. No robotics-specific engineering. If the result is 19x, the barrier to entry for automating general-purpose robotics has just fallen substantially.

We must be honest about what this means in the short term: more pressure on technical integration and commissioning roles in robotics, manufacturing and logistics sectors that are already in accelerated transition. The World Economic Forum confirmed at Davos in January 2026 that the robotic deployment era —entire fleets of robots moving cargo containers without human intervention— is already present, not future. What Project Fetch adds is that even the most demanding phase of that deployment —the initial hardware programming— is being automated at speeds that leave little time for adaptation.

In the long term, however, this kind of capability is exactly the infrastructure that makes the most promising scenario possible: robots capable enough and cheap enough to program that physical automation reaches not only large corporations, but hospitals, medical research labs, home environments. The same general scaling that led Opus 4.7 to program a robot dog in ten minutes is what, extended several more years, could program robotic arms in operating rooms or assistance systems for dependent care. The path between here and there is real and not free of disruption, but the direction is clearly signposted.