Billing & tokens
How billing works in AgentHere: tokens, the free grant, top-up packs, usage, and payments via Robokassa.
AgentHere runs on a prepaid token model: you top up a balance of tokens and they're debited as generation happens. Tokens don't expire. In addition to token packs, there is a subscription to agent plans.
What a token is
A token is a piece of text the model reads or writes (roughly ¾ of a word). Both the request and the reply are metered: input tokens (your prompt + context) and output tokens (the agent's reply). The longer the context and reply, the higher the spend.
The free grant
On sign-up, 1,000,000 tokens are credited once. It's a one-time bonus, not a monthly quota — once spent, you top up. The free tokens are enough to test agents and learn what your tasks cost.
Top-up packs
| Pack | Price | Tokens |
|---|---|---|
| Starter | 200 ₽ | 2,000,000 |
| Plus | 600 ₽ | 7,000,000 |
| Pro | 1,500 ₽ | 18,000,000 |
| Scale | 5,500 ₽ | 70,000,000 |
Current prices and packs are on the pricing page. Payment is one-time, no auto-renewal.
How to top up
- Go to the top-up section from the dashboard.
- Pick a pack.
- Pay by card via Robokassa.
After a successful payment, tokens are credited automatically, and a fiscal receipt (54-FZ) shows up in the payments section.
Usage limits
- No daily cap — overspend is guarded by the balance: spend is debited as it happens, and a zero balance makes requests error out.
- 5 messages per hour for anonymous chat.
- 60 requests per minute per API key.
Where to see spend
The Usage section shows spend by day, model, and source: total tokens broken down into input, output (billed at the 3× rate), tokens read from cache (÷10 rate), and service operations (media, tools). The history shows which surface each request came from and how many tokens were billed, refunds included. Cloud rent counts toward spend but is shown separately: its rows and other service records are hidden from the history by default (toggle «Show cloud rent and service records»). It helps you optimize prompts and pick the right model for the task.
How much a model costs: Flash vs Pro
Token spend depends on the chosen model:
deepseek-v4-flash— the main model: matches Pro on most tasks, answers faster, and consumes fewer tokens.deepseek-v4-pro— for hard multi-step reasoning. From October 5, 2026 Pro consumes 3× more tokens than Flash: the same request on Pro costs three times more against your balance.
We recommend keeping Flash as the default and switching to Pro only for tasks where Flash falls short. Details in the «Flash instead of Pro» post.
How to save tokens
- Shorter system prompt and context — fewer input tokens.
deepseek-v4-flashis cheaper for simple tasks;deepseek-v4-proandglm-5.2cost more but are sharper on hard ones.- Cap
max_tokenswhen you need a short reply.