These animations look fairly breakable.
For example, the horse has a certain degree of separation between it and the rest of the background. You could theoretically compute the difference between one frame and another to see what changed rapidly and what just moved a little bit.
To be honest, these captchas aren't that scary.

What's more is that like many people before my have mentioned, it's too cost-inefficient to produce thousands of unique versions of these.
For example, you would need a unique repository of animations that could be used in order to make unique captcha animations. This repository would have to be huge and secure.
That might now even work.
If you have a 3D model, you just grab two frames and cross-reference it with two frames of a stored animation. You wouldn't even need to have the originals as a team could extrapolate and develop their own 3D models to compare these to.
Before I get on a brainstorming tangent, let me just say that this probably won't happen.
