LLM is spicy memoizing, so it can potentially be faster than a tool call. But people will spend a month tweaking and testing to ensure they have the level of determinism they need, which means it's more expensive, and that they should have used actual memoization in the first place.