On a per-token, it's cheaper than Opus, GPT, and Gemini Pro; and while I hear the "it uses more tokens so its more expensive", this discounts a few things (1) improvements over time (2) finding the right way to prompt it (3) finding proper places to use this model.