Anthropic's Claude 5.2 Leaps in Robot Control, Ex-Tesla Optimus Researcher Joins

Anthropic Releases Claude Fable 5.2 and Mythos 5.2

Anthropic unveiled two new flagship models, Claude Fable 5.2 and Mythos 5.2 . Beyond scientific reasoning — where roughly 1,000 AI agents collaboratively formalized a proof of Fermat's Last Theorem in 30 hours — the release highlights a major advance in robot control.

Robot Control Benchmark Gains

On the RoboDojo Sim benchmark, Fable 5.2 scores 64.9% , up from 27.6% for Fable 5.1 (a 37.3 percentage-point improvement). It also outperforms Claude Opus 5.5 (39.4%) and GPT-6 Astra (29.8%) . Anthropic states that Fable 5.2 exceeds most AI systems built specifically for robot control.

Official demo videos show Fable 5.2 controlling robots to manipulate puzzle blocks, unscrew bottle caps, and draw robot patterns.

The "Bitter Lesson" Extends to Robotics

Anthropic product manager Alex Albert declared that the "bitter lesson" has arrived for robotics. The concept, originating from Rich Sutton's 2019 essay, argues that as compute and model scale grow, general learning methods eventually outperform hand-crafted approaches on a widening range of tasks.

This discussion is now backed by peer-reviewed work. In October 2025, Robocurve and collaborators published "The Bitter Lesson Comes to Robotics" , evaluating frontier general multimodal models on real robot arms. Key findings:

General models can already complete diverse real-world manipulation tasks.

Capabilities have improved steadily over months.

Precision, execution speed, and inference cost remain significant practical constraints.

Paper co-author Jay Chooi echoed the "bitter lesson" framing after the Fable 5.2 release.

Anthropic's Prior Robotics Research

In July 2025, Anthropic published "Claude plays robotics" , testing Claude on simulated robots, robot arms, and a real quadruped. The study explored three control modes:

Generating low-level control commands

Writing control programs

Calling pre-trained robot policies

Results showed Claude possesses meaningful environment understanding and operation planning, though performance depends heavily on control interface, action precision, and inference latency.

Former Tesla Optimus Researcher Joins Anthropic

Bingjie Tang , previously a researcher on Tesla's Optimus humanoid robot, has joined Anthropic as a Member of Technical Staff. Her background aligns closely with Anthropic's robotics push:

Tesla Optimus AI team : Learning-based dexterous manipulation, covering data collection, policy training, evaluation, and real-robot deployment.

PhD, University of Southern California : Collaborated with NVIDIA Seattle Robotics Lab on contact-rich manipulation, robot learning, and Sim-to-Real transfer.

Sim-to-Real challenges : Friction, contact states, object geometry, and hardware errors create gaps between simulation and reality.

Key projects : Refinery, FORGE — addressing policy learning, real-world interaction, and manipulation adaptation.

Personal homepage:

https://bingjietang718.github.io/