Runs and spending
You can see what each answer cost
A run is laid out node by node: what the search found, what went into the model, what it returned, how long it took and what it cost. On the free plan the bill does not grow past three hundred quarks: once they are gone, the bot stops answering until the new month.
| Node | What it returned | Time | Quarks |
|---|---|---|---|
| Customer question | widget · «How much is delivery to Kazan?» | 0.0 s | 0.40 |
| Search the documents | 3 fragments from «Rates and policies» | 0.3 s | 0.00 |
| Answer to the customer | GPT-6 Luna · 1,940 tokens in, 88 out | 1.4 s | 0.34 |
| Send to the customer | to the widget the question came from | 0.0 s | 0.00 |
| Ticket for a manager | not needed: the answer was in the documents | — | — |
A run on 19 August, 14:22. Five rows, a time and a price on each, including the node that was not needed: the answer was in the documents, so no ticket was opened.
300 quarks a month once your address is confirmed, no card needed. What your answer costs depends on the model and on the length of what was found, and the very first run shows it.
Once the bot is answering customers
«What if it answers the wrong thing?»
A bad answer is fixed from the log
The node's card shows what went into the model and what came back: the instruction, the fragments found, the customer's question and the earlier turns from memory.
That is usually where the cause is visible: too much found text travelled into the model, or the instruction says the wrong thing. One paragraph is edited, and the difference shows on the next run. What to do when the bot answers the wrong thing is covered in the documentation.
Next
Where that number came from
Four tenths of a quark for running the scenario and one tenth for each model call, then the length of the text: every model has its own rate per thousand tokens, input and output counted separately. In the example above that is GPT-6 Luna, the model a new node picks by default: 1,940 tokens of question together with the fragments found, and 88 tokens of answer, which is 0.24 quarks. 0.74 in total for this conversation.
You pick the model, and another one gives another number. Web search, an image and speech are counted as their own lines, and all of them are visible in the same breakdown next to their node.
One answer to a customer
- Running the scenario0.40
- The model call0.10
- 1,940 tokens in0.194
- 88 tokens back0.044
- Total0.74
One particular conversation on one model, not a tariff.
Three hundred quarks to see your own numbers
Your documents and your questions show the spending more precisely than any table on this page. No card needed.
See it on your own numbers