Fable 5.2 hits 64.9% on RoboDojo; can frontier models replace dedicated robot policies?

Buffalo4516 · reddit · 2026-11-13

Fable 5.2 listing robot benchmarks on its launch page revives the question: do robots need specialized embodied models?

Fable 5.2 scored 64.9% on Anthropic's own RoboDojo sim (dual-arm robot driven from camera images), more than double Fable 5.1.
An independent study (arXiv 2609.24170) ran zero-shot GPT-6 Astra on RoboDojo, outranking all 40 public policies — yet precision tasks averaged just 4%, and real-robot testing ended after hardware damage.
Plugging a frozen LingBot-Touch encoder into four separate robot policies lifted 26-task averages by 14–25 points, with the largest gains where cameras can't resolve.
The poster's take: the frontier model calls the embodied model like Claude calls tools — it plans, the trained model executes; scaling vision and compute alone may not close the gap.

More from Embodied