Guotai Haitong has released a research report stating that Anthropic's launch of Claude Fable 5.2/Mythos 5.2 further enhances the robot action decision-making capabilities of general-purpose large models, driving general models to penetrate into the action strategy layer of embodied intelligence, with core production segments and supply chain companies with North American operations poised to benefit.
The main views of Guotai Haitong Securities Co., Ltd. (601211) are as follows: Event: Fable 5.2 significantly outperformed Astra in publicly available robot simulation operation tests and has demonstrated real-machine operation capabilities.
On November 9, 2026 local time, U.S.-based Anthropic released Claude Fable 5.2/Mythos 5.2, further enhancing the robot action decision-making capabilities of general-purpose large models and driving general models to penetrate into the action strategy layer of embodied intelligence. In the RoboDojo benchmark test, Fable 5.2 scored 64.9%, approximately 2.2 times that of GPT6-Astra, representing a 37.3 percentage point improvement over the previous generation Fable 5.1 and a 25.5 percentage point improvement over Opus 5.5.
This is important new evidence of general-purpose large models entering the embodied intelligence field. The firm believes that general-purpose foundation models from different companies may all be able to cross the capability threshold for embodied action decision-making, and the pace of improvement is very fast. This will greatly accelerate the commercialization of the embodied intelligence industry.
Previously, Astra had already proven that general-purpose large models can directly output physical action commands without specialized robot fine-tuning. Fable 5.2 once again demonstrates that general-purpose foundation models may gradually absorb the high-level action decision-making capabilities of specialized VLA models, and the boundary between robot models and general models is becoming increasingly blurred. The independent value of specialized VLA is facing reassessment.
The more adept general models become at directly outputting action targets, the more specialized VLA must enhance its advantages in inference cost, real-robot success rate, action frequency, and deployability. General models cannot yet replace the real-time control systems of specialized robots, and the value of the execution layer JEU may become even more prominent.
Large models are responsible for deciding actions, but actual execution still requires stable state estimation, force control, collision response, joint control, and edge-side real-time closed-loop systems. Improved general model capabilities may also expand demand for high-quality execution systems.
Risk warning: Risks of embodied intelligence model technology development falling short of expectations, and risks of downstream demand falling short of expectations.