--- title: "Anthropic's Opus 4.7 Programs a Robodog 20x Faster Than Humans" subtitle: "Claude Opus 4.7 autonomously programmed a robodog 20x faster than the best human team, signaling a new era of physical agentic AI" url: "https://alphasignal.ai/news/anthropic-s-opus-4-7-programs-a-robodog-20x-faster-than-humans" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Anthropic" company_url: "https://alphasignal.ai/company/anthropicai" published: "2026-06-18T16:52:34.000Z" updated: "2026-06-18T17:23:17.557Z" type: "News" topics: - "Robotics" - "Security" subtopics: - "Vla Models" - "Manipulation" - "Red Teaming" upvotes: 2495 read_time: "6 min" source_url: "https://www.anthropic.com/research/project-fetch-phase-two" comments: 0 --- # Anthropic's Opus 4.7 Programs a Robodog 20x Faster Than Humans > Claude Opus 4.7 autonomously programmed a robodog 20x faster than the best human team, signaling a new era of physical agentic AI Anthropic's [Frontier Red Team](https://www.anthropic.com/research/project-fetch-phase-two) just published the results of Project Fetch Phase Two, and the numbers are striking. Less than a year after Claude helped human teams program a quadruped robodog, the latest model did the same tasks entirely on its own , and did them roughly 20 times faster than the fastest human team from the original experiment. This is not a robotics paper. It's a capability benchmark, and what it reveals about the pace of AI progress is worth paying attention to. ## A quick recap of Phase One Back in August 2024, Anthropic ran an internal experiment: two teams of non-expert employees were given a [Unitree Go2 quadruped robodog](https://www.anthropic.com/research/project-fetch-robot-dog) and asked to program it to fetch a beach ball. One team was randomly assigned to use Claude (Team Claude), the other had to rely only on the internet and their own ingenuity (Team Claude-less). Overall, Team Claude accomplished more tasks and completed them faster , for the tasks both teams finished, Team Claude succeeded in about half the time it took Team Claude-less. Before running Phase Two, the team checked whether Opus 4.1 (the model powering Phase One) could handle the tasks entirely on its own. It could not. Like the human team without Claude, it got stuck on the very first step: connecting to the robot. That was less than a year ago. ## What changed in Phase Two For this autonomous update, researchers ran three trials of Opus 4.7 using adaptive thinking with effort set to maximum in Claude Code. Adaptive thinking is Anthropic's term for dynamic compute allocation , the model decides how much reasoning to apply based on task difficulty, spending more tokens on harder problems. The role of the human researcher was limited to plugging a laptop running Claude Code into the robodog, entering the initial prompt, approving commands, and approving the model to move to the next task. The tasks covered the full software side of robot programming: - Connecting to the robodog's video camera - Connecting to the lidar sensor (the sensor the robot uses to map its surroundings) - Writing a program for manual control - Building a path-monitoring system - Detecting the beach ball using computer vision ![Comparison table showing Project Fetch Phase Two task completion times across Team Claude-less, Team Claude, and Opus 4.7 alone](https://www-cdn.anthropic.com/images/4zrzovbb/website/5d1655f4a40044244b6fae0c3826e668a8422bd3-4005x2308.png) ## The numbers On every task completed by at least one human team in August, Opus 4.7 completed the same task at least ten times faster. For the four tasks completed by both human teams, Opus 4.7 was on average more than 37 times faster than Team Claude-less and more than 18 times faster than Team Claude. In raw time: Team Claude-less took 361 minutes, Team Claude took 181 minutes, and Opus 4.7 averaged 9 minutes and 35 seconds. ![Bar chart comparing total completion times: Team Claude-less at 361 min, Team Claude at 181 min, Opus 4.7 at 9:35](https://www-cdn.anthropic.com/images/4zrzovbb/website/0424a5a4ed31f7891e4048d091aefa52f36862cf-1999x1092.png) Code volume tells an equally interesting story. Whereas Team Claude wrote 10,309 lines of code and Team Claude-less wrote 1,136 lines, Opus 4.7 produced just 1,045 lines , achieving the same or better results with almost ten times less code than the Claude-assisted human team. This is a sign of efficiency, not just speed: the model found shorter, cleaner paths to working solutions. Whereas the humans struggled to choose between multiple different approaches to interface with the dog's sensors, Opus 4.7 was able to quickly identify the best path, and much of the code it wrote was effective on the first try. ## Where it still falls short The "fetching" part of Project Fetch remains unsolved. The final task , autonomously nudging a beach ball back to a home base , requires something closer to closed-loop physical control: perceive the ball's position, infer how the last command affected it, adjust the next command accordingly. Claude struggled to capture this subtlety. Like the human participants before it, Claude was able to move the robot behind the ball and position it to knock the ball back to the starting point , but the efforts were poorly controlled and not successful. This is a meaningful distinction. The tasks Opus 4.7 crushed were fundamentally software tasks: parsing APIs, writing sensor integration code, running object detection pipelines. The task it failed was a real-time physical control problem requiring tight feedback between perception and action , something that humans, with a controller in hand, can do after a few minutes of practice. One researcher with more robotics experience than the Phase One volunteers successfully accomplished the task of programming autonomous fetching. With more time and additional scaffolding, Anthropic believes it is very likely that current Claude models could do the same. ## The real signal: general scaling, not robotics fine-tuning Here's the part that should update your priors. This progress is not the result of a concerted effort to improve the robotics capabilities of Anthropic's models. These improvements have emerged from much more general scaling. Nobody trained Claude on robot SDK documentation or sensor fusion tutorials. The capability just showed up. This echoes a broader pattern Anthropic has documented across domains. First, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves. Anthropic has seen this in cybersecurity, and now the same dynamics are starting to take shape at the intersection of AI and the physical world. The parallel to agentic coding is deliberate. Models now complete what was previously pair-programming work between humans and models much more quickly by themselves, which means that people can more quickly transition to controlling and using the robots. Just as Claude Code went from autocomplete to autonomous multi-file refactoring, it may be on a similar trajectory with physical hardware. ## What this means for physical agentic AI The experiment used a [Unitree Go2](https://www.anthropic.com/research/project-fetch-robot-dog), an off-the-shelf quadruped that costs around $16,900 , cheap by robotics standards. The point isn't that Claude can now run a warehouse. The point is that the barrier to interfacing with physical hardware through code is collapsing fast. We are plausibly entering the early era of physical agentic AI , similar to how AI models used existing software editing tools when they made the transition to more agentic coding. The analogy is apt: just as Claude Code doesn't need to understand how a compiler works to write correct code, it doesn't need to understand robot kinematics to write correct sensor integration scripts. Separately, independent research using the Claude Agent SDK has shown that across three standard robotics benchmarks totaling 184 manipulation tasks, a Claude-based agentic approach achieves 84.9–96% success rates with zero demonstrations , matching models that were fine-tuned with 50–100 human-guided examples. The read-reason-write-retry loop that makes Claude Code effective in software appears to transfer directly to robotics simulation. More research is needed to understand models' ability to make physical tools more bespoke, whether by writing control policies tailored to particular tasks or by designing robotic systems. But as we have seen, apparently large distances in model capability can be traversed quickly. It would be unwise to rule out the same trajectory in hardware. ## Sources - [X post by @anthropicai](https://x.com/anthropicai/status/2067651699486200091) - [Project Fetch: Phase Two — Anthropic](https://www.anthropic.com/research/project-fetch-phase-two) - [Project Fetch: Can Claude train a robot dog? — Anthropic](https://www.anthropic.com/research/project-fetch-robot-dog) - [Claude Code As Embodied Agent to Control Robots — Medium](https://medium.com/@brianytsui/claude-code-as-embodied-agent-to-control-robots-85-96-success-in-sim-with-zero-demonstration-455a9044d353) - [Claude Code — Anthropic Product Page](https://www.anthropic.com/product/claude-code) - [Anthropic Project Fetch tests Claude's robotics uplift — EdTech Innovation Hub](https://www.edtechinnovationhub.com/news/claude-shows-speed-advantage-in-anthropics-project-fetch-robot-dog-experiment)