It's pretty good! It did quite well when asked a variant of "when do the trains meet" problem, although it used different values for speed than what I told it to use (and what it actually used in textual response: 120km/h and 20km/h in video vs 80km/h and 60km/h in the prompt and textual response).

https://math-gpt.org/?video_id=a7489ec5-b06e-480a-96c4-27765...

If it did the animations in 3Blue1Brown style, that would be cherry on the top! ;-)

Edit: I did notice it uses Manim, it just doesn't have the same feeling