Skip to main content
A variant changes how a request to a model is handled without changing the model. Add it to the model field:
There are two kinds:
  • Routing variants work on any model id. They are not listed in GET /v1/models; the model’s metadata, context window, and base price stay the same.
  • Catalog variants are separate entries in GET /v1/models with their own id, price, and limits. Use them only on models that list them.

​
Routing variants

​
:nitro

Providers are ordered by median output tokens per second over the last 30 minutes (or the last day when there are too few recent requests). Providers without enough measurements follow in the default order. Use it for long outputs and interactive streaming.

​
:floor

Providers are ordered by price only. The default order also penalizes much slower providers; :floor does not. Providers below 90% uptime still go last. Use it for batch and background jobs where cost matters more than speed.

​
:exacto

Providers are ordered by success rate over the last 30 minutes (or the last day), then by median time to first token, then by the default order. Use it for agents and tool calling, where a failed request costs more than a slower one. :exacto is ignored when the request also sets a sort (:nitro, :floor, or provider.sort).

​
:<effort>

Sets the reasoning effort, like reasoning_effort in the body. model@high is accepted as well.
  • If the suffix and the body reasoning_effort disagree, the request fails with 400.
  • If the model does not support that effort, the request fails with 400 and the message lists the supported values. A body-only reasoning_effort the model does not support is dropped silently instead.

​
:reasoning-exclude

Hides reasoning text from the response. You still pay for the reasoning tokens. See Chat Completions.

​
Prefixes

Provider ids come from GET /v1/models/{model}/providers. An unknown provider id returns 400, code unknown_provider.

​
Catalog variants

Check GET /v1/models for the ids that exist. A catalog variant is part of the model id, so routing variants go after it: gpt-oss-20b:free:nitro.

​
Combining variants

Variants combine in any order:
Rules when variants overlap: Variants are not accepted on auto (auto:nitro returns 400). The Auto Router chooses the effort itself; send reasoning_effort in the body to suggest one, and use the provider object for provider preferences.

​
In-chat image generation

On /v1/chat/completions, @<image-model> after the model id generates an image inside the chat reply, for example gpt-5.1@flux-2. This is not a routing variant. See In-chat Image Generation.