Claude's Fable 5.2 tops robot models on RoboDojo

Anthropic's Fable 5.2 outperformed other models and most AI systems built to control robots on RoboDojo, a benchmark built around a simulated two-arm robot, according to a post from the @claudeai account. The post says the model shows substantially improved skills at controlling physical robots, including on tasks it has never explicitly been trained to do.

Claude

@claudeai

X

it outperformed other models and most AI systems built to control robots

That is the whole claim from the company. No per task scores, no list of which robot policies it was measured against, no detail on how the two-arm setup was driven. If you want a number beyond the ranking, there is not one in the post.

Someone is running it live

Jiafei Duan says Fable 5.2 is being shown in a live head to head comparison against other models at CoRL2026, at the BitRobot booth 408 for RoboColosseum, the open evaluation effort that has been running large batches of real robot trials. That matters more than the announcement. An in-person head to head at a conference booth is the kind of thing where a model either holds up in front of people who build these systems or it does not.

Shown versus claimed

RoboDojo is simulation. A general purpose chat model beating purpose-built robot policies in a simulated two-arm environment is a real result for the benchmark, and it is not the same as beating them on hardware, where contact, latency and sensor noise do most of the damage. Nobody in the sources has shown Fable 5.2 running a physical robot.

The interesting part is the direction. Robot learning has mostly assumed you need a model trained on robot data to control a robot. A frontier language model placing near the top of a manipulation benchmark, on tasks it was not trained for, is a different bet about where the capability comes from. Whether it survives contact with an actual arm is the next question, and right now the answer is at booth 408.

Jiafei Duan✈️CoRL2026

@DJiafei

X

We will be showcasing Fable 5.2 live head to head comparison today at #CoRL2026 BitRobot booth