Quarks and cost
The platform counts in quarks. It is the service’s unit of account: quarks arrive with your plan and are spent on work actually done — on the run happening, on model calls, and on paid external operations. A seat on a plan does not consume quarks by itself; money is what you pay for the plan or a pack.
Every model has its own rate in quarks per thousand tokens, input and output counted
separately, and it is shown next to the model’s name in the list. For gpt-6-luna that is
0.1 in and 0.5 out — an inexpensive model answers for less than a strong one, and the
difference is visible before the run.
The formula
One run costs:
0.4 (the run itself)
+ per LLM node: 0.1 + input/1000 × k_in + output/1000 × k_out
+ media, web search, embeddings and reranking — as their own linesk_in and k_out are that model’s rate per thousand tokens, input and output counted
separately. The rate is declared up front and does not move from run to run, so a turn’s
price can be worked out before you make it.
A typical chat turn on the default model — 1,940 tokens of question and found fragments,
88 tokens of answer — costs 0.4 + 0.1 + 1.94 × 0.1 + 0.088 × 0.5 = 0.74 quarks, about
three quarters of a quark. The rates
are fractional on purpose: rounding the overheads up to a whole quark would restore the
very skew the scale was changed to remove.
An unknown model does not become free: it falls back to a default rate. A zero in the catalog means the vendor gives the model away for free, not that its price is unknown — those are different things, and confusing them is expensive.
Billed separately
| What | How it is counted |
|---|---|
| Scenario run | Fixed, once per run |
| LLM node call | Fixed per call + tokens at the model’s rate |
| Image and speech generation | By what the generation itself cost; a flat rate when that is not known |
| Web search | By the search cost when it is visible separately; a flat rate otherwise |
| Embeddings and reranking | By tokens |
| Code node | Fixed per code run |
Where to see the spend
The wallet at the bottom of the sidebar shows what is left; clicking it opens the spend screen with a breakdown by day, scenario and channel. Every run in Executions shows the tokens and quarks it cost and which node spent them.
The assistant’s debug runs are not billed: the model decides to make them, not you.
What next
- Plan limits — what each of the four tiers includes.
- Models — what models are available and where their rate is shown.
- Your own models and keys — running on a key of your own.