China's relay market: what a million tokens costs

Discount trend

Market median effective price ÷ official reference. Flagship models sit around 15% because most sites sell "$1 of quota for ¥1" while copying official list prices; the fast tier trades near official. A measurement, not a recommendation of any tier.

Task cost: how many tokens each model spends on the same job

Tokens carry three bills: production cost, market unit price and task cost. This page gives the first two; the third comes from the same 30 short tasks we send every model daily: record the output tokens each task actually consumed, multiply by the unit price, and you get "what this job costs".

Definitions, data and citation

Per model per day: cross-site median effective price per million output tokens (≥15 sites); discount = median effective ÷ official reference; tier and whole-market discount and median take the median across models (a daily snapshot, affected by models entering or leaving); the level starts at 2026-09-02 = 100 and chains the geometric mean of day-over-day changes in the medians of models present on both days (unaffected by composition, used for direction); history is rebuilt day by day from quote validity intervals.

Cost layer: GPU rental ledger (what one GPU costs per hour).

Data: price-index.json (full series, updated daily) · cite as "Sinan Token Price Index, compute.sinanlab.com".

<a href="https://compute.sinanlab.com/en/price-index"><img src="https://compute.sinanlab.com/badge/price-index.svg" alt="Sinan Token Price Index" width="360" height="72"></a>

Governance rules (definitions, release cadence, change process, conflicts of interest):/governance

Cite this page

Data may be freely cited and republished with attribution and a link. We take no money from measured channels, sell no rankings and add no referral parameters, so you can cite us as an independent third-party measurement.

Media, research institutions and financial institutions needing raw data or custom definitions: write to [email protected], see Press & researchers。

Do you want Sinan to route your calls for you?

The idea: you hold your own Keys for each site; Sinan only measures, filters by your criteria, routes requests to the channel you chose, and logs latency and cost. Right now weNot doneVotes decide whether this gets built.