Anthropic Unveils Claude Fable 5.2 with Major Robotics Control Gains

Anthropic Is Serious About Robotics

Machine Heart Editorial Team

In the early hours of this morning, Anthropic released two top-tier models at once: Claude Fable 5.2 and Mythos 5.2. The new models demonstrated multiple achievements in scientific research, including completing the formalization of the proof of Fermat's Last Theorem in just 30 hours through the collaboration of approximately 1,000 AI agents.

In this update, robot control capabilities have also attracted considerable attention. In the official release article, the science and engineering section specifically includes a "Hardware and Robotics" subsection.

In the official demo video, Claude Fable 5.2 controls a robot to complete tasks such as manipulating puzzle blocks, unscrewing bottle caps, and drawing robot patterns.

According to data published by Anthropic, Claude Fable 5.2 achieved a score of 64.9% on RoboDojo Sim, an improvement of 37.3 percentage points over the previous generation Fable 5.1's 27.6%, and also surpassing Claude Opus 5.5 (39.4%) and GPT-6 Astra (29.8%).

Anthropic also stated that Fable 5.2 outperformed most AI systems specifically developed for robot control in this evaluation.

Anthropic Product Manager Alex Albert stated bluntly that the "bitter lesson" for robotics has arrived.

Meanwhile, researcher Bingjie Tang, who previously participated in the development of Tesla's Optimus humanoid robot, has also joined Anthropic. She previously focused on dexterous robotic manipulation, policy learning, and real-world deployment, and her research background forms a noteworthy intersection with the robotic capabilities recently demonstrated by Anthropic.

Anthropic Is Expanding Claude's Robot Control Capabilities

In fact, Anthropic has publicly released robotics-related research multiple times this year.

In July, Anthropic published "Claude plays robotics," testing Claude's performance on simulated robots, robotic arms, and real quadruped robots.

Researchers tried various control approaches, including generating low-level control commands, writing control programs, and invoking pretrained robot policies.

Experiments found that Claude already possesses a certain degree of robot environment understanding and operation planning capability, with specific performance affected by factors such as control interfaces, motion precision, and inference latency.

By October, research institutions such as Robocurve published the paper "The Bitter Lesson Comes to Robotics," further evaluating the ability of frontier general-purpose multimodal models to control real robotic arms.

The study found that general-purpose models are already capable of completing various physical robot manipulation tasks, with related capabilities continuing to improve over several months. At the same time, manipulation precision, execution speed, and inference cost remain important constraints in practical applications.

Jay Chooi, one of the paper's authors, also mentioned the "bitter lesson" in the robotics field on social media after the release of Fable 5.2.

The "bitter lesson" originates from a classic viewpoint in the machine learning field: as computational resources and model scale grow, general learning methods tend to become competitive on an increasing number of tasks.

Today, this discussion is also beginning to extend into the robotics field.

As the visual understanding, spatial reasoning, and task planning capabilities of general-purpose models continue to improve, researchers are exploring how much work they can take on within robot control systems.

Former Tesla Optimus Researcher Joins Anthropic

Just as Anthropic demonstrated its robotic capabilities, news of a talent move also drew attention.

Researcher Bingjie Tang, who previously participated in the development of Tesla's Optimus humanoid robot, recently confirmed joining Anthropic as a Member of Technical Staff.

She previously worked on learning-based dexterous manipulation research in the Tesla Optimus AI team, with work covering data collection, robot policy training, evaluation, and real robot deployment.

Before joining Tesla, Bingjie Tang earned a PhD in Computer Science from the University of Southern California. During her doctoral studies, she worked closely with researchers at the NVIDIA Seattle Robotics Lab, conducting research on contact-rich robotic manipulation, robot learning, and Sim-to-Real.

Among these, Sim-to-Real is one of the important directions in current robot learning, primarily studying how to transfer manipulation policies trained in simulation environments to real robots.

Friction, contact states, object shapes, and hardware errors in real environments all affect a robot's motion performance. Therefore, from simulation training to real-world deployment, there remain numerous technical challenges to be solved.

Bingjie Tang has previously participated in research such as Refinery and FORGE, which also involve robot policy learning, real-world interaction, and manipulation adaptation.

Looking at her personal track record, she has research experience spanning data, algorithm training, and real-world robot deployment, which has strong technical relevance to the robotic capabilities recently demonstrated by Anthropic.

In 2026, embodied intelligence continues to heat up. From world models and spatial intelligence to robot foundation models, more and more AI companies are beginning to explore the physical world, while the robotic control capabilities of general-purpose large models are also advancing rapidly.

As the capabilities of general-purpose large models continue to improve, what role will they play in future robotic intelligence systems? What new progress will Anthropic bring next? It is worth looking forward to.

© THE END

For reprinting, please contact this official account for authorization

Submissions or story tips: [email protected]

Open in WeChat