You don't need to Ocr. Llms can respond directly to the scanned image. They are better than most Ocr programs.

Indeed the token cost of image inputs are lower because you have more fine grained control of the latent token space