Auto Mode
Virtual model IDs that pick the model for you, within a price band you choose.
Auto Mode gives you three virtual model IDs that delegate the choice of model to a router instead of naming one. You pick a price band; the router picks a model inside it for each request.
The IDs
| Model ID | Band | Max prompt ($/1M) | Max completion ($/1M) |
|---|---|---|---|
auto-cheap | Cheap | 0.25 | 1 |
auto | Balanced | 1.50 | 6 |
auto-quality | Quality | 10 | 50 |
Use them exactly like any other model ID:
model: 'auto'What the router can pick
The candidate pool is derived from the live catalog on every request, not from a hardcoded list, so new models join it and superseded ones drop out with no change on your side. A model is eligible when it:
- comes from an approved vendor (Anthropic, OpenAI, Google, DeepSeek, Qwen, xAI, Moonshot, Z.ai, Meta, Mistral),
- supports tool calling,
- was first listed within the last 12 months,
- and prices inside the band above.
Preview, experimental, and :free/:batch variants are never routed to.
Pricing
The band's caps are the published price for the virtual ID — that's what's reserved before the call, and it's the ceiling the upstream router is held to, so you're never billed above it. Where the upstream reports the real cost of the model it picked, that's what you're charged, so the usual case is less than the published rate. Everything else about billing is unchanged.
Knowing what you got
Auto Mode responses carry three extra headers:
| Header | Meaning |
|---|---|
x-token-kiosk-routed-model | The model that actually served the request |
x-token-kiosk-router-tier | cheap, balanced, or quality |
x-token-kiosk-router-latency-ms | Upstream latency for the call |
On streaming responses these are sent before the first chunk.