Former Tesla Optimus Engineer Joins Anthropic as Claude 5.2 Sets Robot Benchmark Record

Background: Bingjie Tang's Career

Bingjie Tang completed her undergraduate studies at Huazhong University of Science and Technology, earned a master's in computer science from Brown University, and received a Ph.D. in computer science from the University of Southern California under advisor Gaurav S. Sukhatme.

Research at NVIDIA Seattle Robotics Lab (2022–2024)

During her Ph.D., Tang interned at NVIDIA's Seattle Robotics Lab, collaborating with researchers including Yashraj Narang and Dieter Fox. Her work focused on sim-to-real transfer for robotic assembly:

IndustReal (RSS 2023, co-first author): The team trained a Franka arm in simulation on peg insertion and gear assembly, then deployed policies to the real robot. They discovered that simulation sometimes produced interpenetration artifacts; downweighting those experiences in training prevented the robot from learning infeasible actions.

AutoMate (RSS 2024, first author): Extended the approach to 100 diverse assembly tasks. Specialist policies trained per task achieved ~80% or higher success rates on 80 of the 100 tasks in simulation.

SRSA (ICLR 2025 Spotlight): Investigated how to select and adapt existing skills for new tasks.

Refinery (first author): Explored how a deployed manipulation policy can be refined using real-world interaction feedback.

Tesla Optimus

After her Ph.D., Tang joined Tesla's Optimus AI team as a Senior Robot Machine Learning Engineer. According to her personal page, she was responsible for learning-based dexterous manipulation, covering data collection, policy training, evaluation, and real-robot deployment.

Anthropic Releases Claude Fable 5.2 and Mythos 5.2

On November 10, the same day Tang announced her move, Anthropic launched two new models sharing a common base: Claude Fable 5.2 (broadly available) and Claude Mythos 5.2 (available through security, life-science verification programs and enterprise products). Both target complex coding, research, and engineering tasks. Compared to the previous generation, output speed increased up to 30%, and standard API pricing dropped to $8 per million input tokens and $40 per million output tokens.

Benchmark Results

Robotics: Fable 5.2 scored 64.9% on a bimanual robot simulation benchmark, a 25.5 percentage-point improvement over Opus 5.5.

Coding: On FrontierSWE v2, Fable 5.2 achieved 69.9%, surpassing Opus 5.5 (62.3%) and Fable 5.1 (56.3%).

Formal theorem proving: An internal research model previously took 11 days to formalize the proof of Fermat's Last Theorem in Lean. With ~1,000 agents running an early Fable 5.2, the same task completed in ~30 hours.

Advanced Capabilities Demonstrated

Electronic design (with Diode): Fable 5.2 designed three complete devices—a handheld electronic pet, a home voice assistant, and a USB conference speaker—producing PCB layouts, enclosures, and firmware. Diode engineers handled manufacturing and assembly. Earlier models struggled with stable PCB layout and component fit.

Robot control: Videos show robots under Fable 5.2 control solving a Rubik's cube, unscrewing bottle caps, and drawing a robot portrait. The article highlights the generalization challenge: opening a bottle when its position or cap size changes, a problem Tang has worked on since her NVIDIA assembly research and Tesla Optimus days.

Significance

Tang's move brings deep expertise in closing the sim-to-real gap for dexterous manipulation directly into Anthropic's efforts to ground large language models in physical robotics. Her background aligns with the new models' demonstrated ability to control robots that must adapt to varying conditions—a key step toward general-purpose robotic foundation models.