the key thing to realize is the Haiku tokenizer consumes 1.5x the tokens for the same text compared to Luna.

So at minimum costing 1.5x for the same request and if input tokens cross 100k then all tokens cost 5x more