Odaily reported that Anthropic has published the second-phase results of Project Fetch, an experiment designed to evaluate how its latest model performs in real-world robotic operations. Rather than limiting the assessment to software workflows or simulated environments, the experiment placed the model in a setting where physical robots had to carry out tasks, giving Anthropic a way to measure how a general-purpose model behaves when interacting with physical tools.
Non-robotics experts used off-the-shelf quadruped robots
According to Anthropic, the second phase of Project Fetch was conducted in August 2025. Researchers asked Anthropic employees who were not robotics experts to use off-the-shelf quadruped robots to complete a series of complex tasks. The experiment compared two approaches: completing tasks with assistance from Claude models, and completing tasks by relying only on human participants and internet resources.
This setup placed the comparison on task execution and model assistance rather than on the specialized background of robotics professionals. The participants were not described as robotics experts, and the hardware used in the experiment consisted of readily available quadruped robots. Anthropic presented the results as evidence of expanded capability in real robotic operations, not as the outcome of training a model specifically for robotics.
Claude Opus 4.7 showed a major speed advantage in autonomous operation
The results showed that, when the latest model Claude Opus 4.7 was allowed to run fully autonomously, it was significantly faster on average than the human teams across all tasks that could be completed. Anthropic said the execution speed improved by at least 10 times. The finding focuses on the model’s ability to move the task process forward directly, rather than only providing step-by-step assistance to a human operator.
Anthropic also stated that this progress was not the result of robotics-specific training. Instead, the company described it as a result of the broader expansion of general large language model capabilities. In Anthropic’s framing, AI is moving from a phase in which it helps humans use tools toward an early stage in which it directly operates physical tools. The company compared this direction with the earlier evolution in software engineering, where AI moved from assisting human users toward agentic programming workflows.