Choosing a model

Name a model, pin a provider with provider/model, call your own models, or send auto. What each form means and the errors you can get.

The model field takes one of four forms.

You send	What happens
`gpt-6-luna`	That model, on whichever of its providers routing ranks first. If that provider fails, the next one serving the same model takes over.
`openai/gpt-6-luna`	That model, only on that provider.
`auto`, `auto:economy`, `auto:balanced`, `auto:best`	DigitSpace chooses the model as well. See `auto`.
`my-vllm/llama-3.3-70b`	A model you added on your own provider key. See Custom endpoints.

Model IDs and aliases are listed by GET /v1/models and on the models page.

Pin a provider

Put the provider's ID in front of the model: anthropic/claude-haiku-5-5, google-ai-studio/gemini-3.6-flash. Routing then only uses that provider's offers of the model. Failover still works between that provider's offers (in different regions, for example) but never moves to another provider.

Pin a provider when you need something only that provider has: a feature, a data-handling term you've agreed with it, or your own key's quota. Otherwise leave it unpinned so failover has somewhere to go.

A DevPass key can't pin a provider; it gets 403 devpass_provider_pinned. Send the model ID alone.

When the model can't be used

Error	Why
`404 model_not_found`	No model has that ID or alias. Check the spelling against `/v1/models`.
`400 capability_mismatch`	No offer of the model can do what the request asks (images, tools, a JSON schema, reasoning, streaming, or that much context).
`403 model_not_allowed`	The key's access rules don't allow the model or its providers.
`503 no_provider_available`	The model exists, but none of its providers can take the request now: they're paused after failures, or the project has no key for them.

For context, routing estimates the request's size as about four characters per token, plus 85 tokens per image, plus max_tokens (512 when not set), and skips offers whose context window is smaller.

Checked against the gateway on 31 October 2026. Something wrong or missing? Tell us.