Avoid reasoning models in any situation where you have low tokens/second