Methodology

Cost per million tokens = price per GPU-hour ÷ tokens produced per GPU-hour. The numerator is built up from hardware, facilities, power and financing; the denominator is set by model and inference-software efficiency.

The formula covers LLM inference. Training compute is treated as capital investment plus amortisation; image, video and speech, billed per image or per second, are handled separately.

Every number on the site carries a date and a source; anything without a direct source is marked as an estimate with its working shown. The build checks every row: data missing a date or source does not go live.

A: exchanges, regulators and official vendor APIs. B: public web pages such as price lists. D: manual compilation and analyst estimates. Paid-licence data (tier C) is never republished; only our own derived results are shown.

Judgement: our analytical conclusion. Speculation: inference without direct evidence. Unverified: from public reporting, not yet checked line by line. Verified: checked against a primary source. Estimate: calculated with the stated method.

If you spot an error, write to us. Once confirmed, we update the data and say publicly what changed, in the dataset changelog and the corrections log.