My understanding is they make a loss overall due to the spending on training new models, that the API costs are profit making if considered in isolation. That said, this is based on guestimates based on hosting costs of open-weight models, owing to a lack of financial transparancey everywhere for the secret-weights models.