June 20, 2026
Bad dog or robot overlord?
Project Fetch: Phase Two
AI robodog got way faster and the comments instantly went full Terminator panic
TLDR: Anthropic says its latest AI handled robodog tasks far faster than last year’s human teams, hinting that AI is getting better at real-world machine work. The comments split hard between impressed, skeptical, and jokingly terrified, with cries of "ad," "rerun it fairly," and "hello, Terminator."
Anthropic came back with Project Fetch: Phase Two, and the headline-grabber is wild: its newer Claude model reportedly handled robodog setup tasks dramatically faster than the human teams from last year—sometimes by double digits, sometimes by a lot more. In plain English, the company says the AI went from "helpful sidekick" to something much closer to "let me do it myself," at least for parts of the job like connecting to sensors, writing code, and getting the machine moving. It still wasn’t perfect at the actual ball-fetching finale, which is exactly the detail skeptics in the comments pounced on.
And oh, the comments did not come to play. One camp basically yelled, "cool demo, still sounds like an ad", with critics arguing this is more of a before-and-after comparison between Anthropic’s old and new model than a true human-versus-AI showdown. Another wanted the real rematch: give today’s humans Claude 4.7 too, then see who wins. Meanwhile, the doomers arrived on schedule with the most clickable reaction of all: "Do you want Terminators?" On the comedy side, someone deadpanned "stop trying to make fetch happen," instantly turning the whole experiment into a Mean Girls meme. Even a wonky line about another model not making an "apples-to-apples comparison" triggered detective mode, with readers speculating about latency, serving setup, and what Anthropic wasn’t spelling out. So yes, the robodog ran—but the real sprint was in the replies.
Key Points
- •Anthropic revisited its August 2024 Project Fetch experiment to test whether newer Claude models could autonomously perform robodog-related tasks better than the previous generation.
- •In the original experiment, non-robotics Anthropic employees using Claude Opus 4.1 outperformed a team without Claude, while Opus 4.1 itself could not complete the tasks independently.
- •In Phase Two, Claude Opus 4.7 reportedly completed every task finished by at least one human team at least 10 times faster, and was about 20 times faster than the fastest human team across completed tasks.
- •Anthropic states that Opus 4.7 still struggled with precise beach-ball movement and that the experiment did not address low-level robotic control problems such as actuation policy development.
- •Phase Two used three trials in Claude Code with adaptive thinking set to maximum, while researcher involvement was limited to setup, prompt entry, command approval, and permission to advance between tasks.