The article repeatedly uses the API charge to the 'customer' as an estimate of the cost value to Anthropic/OpenAI.

Is there any good data on the true cost of inference for these companies?