Claude Fable 5.2 scores 64.9% on RoboDojo, more than double GPT-6 Astra on Anthropic's own run

Anthropic released Claude Fable 5.2 on Nov. 9, a general model pitched at coding, finance and the sciences, and put a robotics result in the launch post. On RoboDojo, a public benchmark in which a model steers a simulated two-arm robot from camera images, Fable 5.2 averaged a 64.9% success rate, Anthropic says. That is more than double GPT-6 Astra at 29.8% and Fable 5.1 at 27.6%, and ahead of Claude Opus 5.5 at 39.4% and Gemini 4 Argon at 34.2%. Anthropic ran every model itself, each at its highest effort setting.

The leap is real on Anthropic's harness, but the order of the rivals behind it depends on who holds the stopwatch. Artificial Analysis's AA-Robotics v0.1, published Oct. 14, drove a real six-axis arm through eight tabletop tasks. There GPT-6 Astra scored 69.4 out of 100 and Fable 5.1 46.3, a gap of 23 points. On Anthropic's RoboDojo run, the same two models sit 2.2 points apart.

Trained robot policies set a low bar in July

RoboDojo's authors released the benchmark in July with 42 simulation tasks on an ARX X5 two-arm platform, testing generalization, memory, precision, long-horizon work and open-ended instructions. Of the 30 trained robot policies they evaluated, the best reached an 8.80% average success rate. Human experts reached 76.03% under the same protocol. Anthropic says Fable 5.2 beat most AI systems built specifically to control robots, which leaves room for a few on the live leaderboard to rank above it. Its figure comes from its own harness, so it sits beside the paper's numbers rather than on the same scale.

Off the benchmark, a puzzle cube and a bottle cap

Anthropic says Fable 5.2 is substantially better at controlling physical robots than earlier Claude models, including on tasks it was never explicitly trained for. Its video shows robots under the model's control solving a puzzle cube, unscrewing a bottle cap and drawing a picture of a robot. In the same post, Anthropic says the model designed three electronic devices with Diode, a circuit-board maker: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. All three worked without electrical modifications.

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1.

RoboDojo scores whether the robot finishes, and a production line pays for how fast it does. Artificial Analysis timed frontier models at about 30 minutes per task against 20 seconds for a person, and Fable 5.1 spent about ten minutes planning before its first move.