Models & availability

Choose an API edition

A catalog family describes an upstream model. An API edition is a specific model definition enabled in Router, with its own enforced limits. Use the exact edition ID in your request.

Check availability

List API models

Authenticated SDK clients can use GET /v1/models. The response includes configured editions even when they have no online providers.

1curl "$ROUTER_BASE_URL/models" \2  -H "Authorization: Bearer $ROUTER_API_KEY"

Each edition includes its exact ID, configured artifact digest, context and output ceilings, prompt-byte limit, capabilities, and observed ready and online provider counts. See the models endpoint reference for every field.

Respect model limits

Use the API edition’s limits. An upstream family’s context window may be larger than Router’s supported context. The current chat API accepts text; catalog metadata about vision or embeddings does not enable those modalities.

Router’s current platform ceilings are 32,768 context tokens, 2,048 output tokens, 16 messages, and 196,608 UTF-8 bytes across all message text. An edition’s limits never exceed its upstream model’s native context, so many editions are lower. Providers can serve a shorter context while they unlock longer ones, so a request is routed only to a provider whose served context fits it: its prompt bytes, its answer limit, and an estimate of its prompt tokens with that answer must all fit what the provider serves. A long prompt therefore goes to a provider serving a longer context, never to one that would refuse it. ready_providers counts only providers that serve the full edition size and answer limit. When no online provider serves a request this long, Router returns 503 no_provider_for_length with Retry-After: 60 rather than shortening the input; send a shorter request or retry later. If a provider that fits is online but busy, you get the normal busy error instead.

Build a model picker

Display the edition title and owner, but send the exact id. Show online providers and free capacity separately. Disable or annotate an edition with no ready providers, while keeping it visible so the user understands that it is supported but temporarily unavailable.

Refresh before a user starts a request, then still handle no_priced_provider_available: discovery is an observation and does not reserve capacity.

After choosing an edition, use Auto or choose a public provider offer. Public offer discovery is separate from total model capacity: providers do not need to publish an offer to remain eligible for Auto.

Browse models in Router.