For one, you don't get to see any output at all if you run out of tokens during thinking.

If you set a limit, once it's hit you just get a failed request with no introspection on where and why CoT went off the rails