We ran a token model for some time for precisely this reason, but came to the realization that the costs were comparatively low. We use OpenAI in the background.