# Service modes

> Pick fast, flex or priority processing with a model:mode suffix, on the providers that offer it.

URL: https://boelabs.com/en/docs/ai-gateway/service-modes
Language: English

A service mode is the same model served differently: faster, cheaper and slower, or with reserved
capacity. Pick one by adding a suffix to the model, the way OpenRouter does:

```json
{ "model": "gpt-6.1-sol:flex", "input": "Summarize this report." }
```

The suffix works on every text contract (`/v1/chat/completions`, `/v1/responses`, the Responses
WebSocket and `/v1/messages`) and on [`/v1/audio/speech`](https://boelabs.com/docs/ai-gateway/speech), and a key scoped to `model`
may also use `model:mode`.

| Suffix | What it is | OpenAI | Anthropic | Gemini |
|---|---|---|---|---|
| `:fast` | Faster output at a premium | `service_tier: "fast"` | `speed: "fast"` | — |
| `:priority` | Reserved capacity at a premium | `service_tier: "priority"` (the same tier as Fast since OpenAI renamed it) | — | `service_tier: "priority"` |
| `:flex` | Cheaper, slower, may be unavailable | `service_tier: "flex"` | — | `service_tier: "flex"` |

## Which models support a mode

A deployment can serve a mode when its provider offers it **and** the catalog prices that tier
(`pricing.serviceTiers`). The gateway routes `model:mode` only to deployments that qualify. When none
does (the model or provider has no such mode, or the deployment goes through an aggregator that does
not expose it), the request is refused **before any upstream call**:

```json
{
  "error": {
    "type": "invalid_request_error",
    "code": "unsupported_service_mode",
    "param": "model",
    "message": "The model `my-model` does not support the `fast` service mode."
  }
}
```

There is no silent fallback to standard processing. A suffix that is not a mode (`:free`, `:online`) is
part of the model's own name.

If the request also sets `service_tier`, it must agree with the suffix (`auto` and `default` defer to
it); otherwise the request fails with `conflicting_service_tier`.

## Billing

The gateway charges the tier the provider says it **served**, not the one requested:

- **OpenAI** reports it in `service_tier`. Fast mode held back by its ramp-rate limit comes back as
  `default` and is charged at standard rates; GPT-5.6 and earlier report Fast as `priority`.
- **Anthropic** reports `usage.speed`; only `fast` is charged at the fast rate.
- **Gemini** Flex never falls back (it fails with 429 or 503 instead), so a Flex answer is charged at
  Flex. Priority overflow is served at Standard, and Gemini says so only in the `x-gemini-service-tier`
  response header; the gateway charges Standard when that header says `standard`.

See [Cost accounting](https://boelabs.com/docs/ai-gateway/cost-accounting) for how tier rates apply.

