# Limits

> Concurrent requests, per-key spend limits, balance reserve and request body size.

| Limit | Value | When exceeded |
| --- | --- | --- |
| Concurrent requests | Limited per account, across all keys | `429 rate_limit_exceeded`, `Retry-After: 1` |
| The model provider's rate limit | Depends on the provider | `429 rate_limit_exceeded`, `Retry-After` when the provider named a pause |
| Key spend limit | Set per key in the cabinet | `403 spend_limit_exceeded` |
| Balance | Free remainder minus the reserve of requests in progress | `402 insufficient_quota` |
| Request body: text | 25 MiB — `/v1/chat/completions`, `/v1/messages`, `/v1/responses` | `413` |
| Request body: media | 70 MiB — `/v1/images/generations`, `/v1/videos` | `413` |
| Images per request | `n` from 1 to 10 | The value is clamped to the range |

## Concurrent requests

The number of requests running at the same time for an account is limited —
across all keys. There is no separate requests-per-minute limit. Once exceeded,
the next request gets `429 rate_limit_exceeded` with a `Retry-After` header —
retry after the indicated pause, when one of the running requests finishes.
Video jobs (`POST /v1/videos`) are not subject to this limit.

If the model's provider hits its rate limit, the gateway also responds `429`;
`Retry-After` is present when the provider named a pause (1 to 60 seconds).

## Spend limit and balance reserve

- A key can have a spend limit: once reached, new requests on that key get
  `403 spend_limit_exceeded` while your other keys keep working.
- While a request runs, its expected cost is reserved on your balance and in
  the key's limit (video jobs hold no reserve). If the remainder is held by requests in progress, a new one
  gets `402` or `403` — retry once they finish.

More on charges in [Balance and pricing](https://ai-seller.vibe-codes.ru/en/docs/pricing).

## Request body size

25 MiB for text endpoints and 70 MiB for images and video — the ceiling for
base64 media attachments. Anything larger gets `413`. Pass large references as
links rather than `data:` URLs when the model supports it.