Are OpenAI and Anthropic Really Losing Money on Inference?

DeepSeek R1 running on H100s puts input-token costs near $0.003 per million—while output tokens still punch in north of $3. That’s a 1,000x spread.

So if a job leans heavy on input—think code linting or parsing big docs—those margins stay fat, even with cautious compute.

System shift: This lop-sided token pricing is bending inference economics. Models that chew on more input and say less? Suddenly the smart bet.