Pricing
What a million tokens costs on each backend, and what you can expect a bill to actually come to.
You pay what the tokens cost. Nothing is marked up at the moment of a request: our fee is taken once, when you add credit, and from then on your balance is drawn down at the real price of the inference you spend. There is no subscription, no minimum, and no per-request commission waiting in the bill.
Which price that is depends on which backend served the request. Tokens produced by the Gonka network are charged at the network's rates; tokens that fell back to OpenRouter are charged at OpenRouter's. It is not a choice you make — the router makes it, per request, and the rate that applied is recorded against the request in your usage log.
The raw rates
What a million tokens costs at each backend, in US dollars:
| Model | Gonka network | OpenRouter fallback | ||
|---|---|---|---|---|
| Input | Output | Input | Output | |
deepseek-v4-flash | $0.001789 | $0.001789 | $0.1477 | $0.2954 |
kimi-k2.6 | $0.0002975 | $0.0002975 | $0.7174 | $3.598 |
minimax-m2.7 | $0.0002975 | $0.0002975 | $0.2943 | $1.266 |
Two things worth reading into that table.
The network charges one rate. On Gonka, input and output are the same number, so a request there costs its total token count times a single price. OpenRouter prices the two separately, with output several times the dearer — which is why the mix of prompt and completion in your traffic only starts to matter once a request falls back.
The OpenRouter figures already include OpenRouter's own fee. Their price list quotes the model rate alone; buying the credit that pays for it carries a payment fee that their documentation leaves out of the quoted number. We pay that fee on the credit serving your fallback requests, and it is folded into the rates above — so these are what those tokens actually cost, not what a price list says they cost. See the note on fallback fees.
Rates are not fixed, and neither table is a contract
Gonka's rates are set by network governance and polled periodically; OpenRouter's move when the upstream provider's do. The rates in force when your request starts are frozen for the whole of it, so a change mid-request cannot alter what that request costs — see Pricing mechanics.
What you will actually pay
The table above prices the two backends. It does not tell you the one thing you came here for, which is what a million tokens costs you — and that depends on how much of your traffic each backend ends up carrying.
So here is the same thing blended: each model's rate weighted by the share of requests each backend actually carried over the last seven days. The network serves the great majority of them, which is why these land close to its column above.
| Model | Input | Output |
|---|---|---|
deepseek-v4-flash | $0.04590 | $0.09056 |
kimi-k2.6 | $0.2171 | $1.088 |
minimax-m2.7 | $0.08920 | $0.3830 |
Treat these as orientation rather than a quote. The split between backends is a property of the network, not of your account, and your own mix will not be exactly the average one: requests that lean harder than most on image input or web search are not priced like plain text, and a workload heavier in them will land above the table.
That said, the figures are measured rather than modelled. They come from every request across all of our users and from the average state of the network over the last seven days, and they are refreshed as both move. As a number to plan a budget around, you can lean on them.
Related
- Pricing endpoint — the same figures as JSON, public and unauthenticated, for a client that wants to price a request itself.
- Pricing mechanics — how the charge for a single request is computed, held and settled.
- Billing — what all this does to your balance, and what a failed request costs.
- The GPU network — why the first columns of that first table are as small as they are.
- Models — what is served today, and which one to reach for.