We've done a bunch of work to strip down the context and minimise the output tokens (which tends to be 100x as slow as input tokens). GPT-4o is pretty fast too :)