Compare

One model, three supplies in one coordinate system: official API, relay API, subscription account or top-up. The first two are priced per million tokens, the third per month — comparing them needs a monthly-usage assumption, and that assumption belongs next to the number, not hidden.

The last column is the cheapest subscription account per month. It is not converted into $/M tokens, because that conversion needs a number no vendor publishes: how many tokens a subscription actually grants. Converting anyway would mean inventing a monthly-usage assumption and hiding it inside a precise-looking figure — worse than showing both units side by side.

ModelOfficial API $/MRelay low $/MRelay median $/MLow / officialCheapest $/mo
GPT 5.6 Sol5.000.1501.000.03×0.590
GPT 5.6 Terra2.000.0600.4000.03×0.590
GPT 5.55.000.1501.000.03×0.590
Opus 55.000.3002.500.06×19.75
Fable 510.000.6003.150.06×19.75
Opus 4.85.000.3002.500.06×19.75
Sonnet 52.000.1201.000.06×19.75
Haiku 4.51.000.0601.000.06×19.75
GPT 5.6 Luna0.2000.0140.2000.07×0.590
Opus 4.75.000.6005.000.12×19.75
Opus 4.65.000.6005.000.12×19.75
Sonnet 4.63.000.3603.000.12×19.75

Official prices are each vendor's published input price, in USD, with no conversion. Relay prices come from model_ratio × group_ratio × $2/M (new-api's 1 ratio = $0.002/1K); groups on expression-based billing are excluded rather than guessed.