> API inference margins are greater than 10% for OpenAI and Anthropic

How do you measure that?

With Amodei's special accounting ofc