GET /v1/models/{model}. The API reference describes each route. Endpoint coverage does not mean every model, optional parameter combination, or production deployment has been live-tested.
The hosted site uses sandbox credits and Stripe test payments, with real inference supplied by the configured upstream provider. The Decisions history UI is available, but the verified provider decision feed remains disconnected. Automatic decision webhooks therefore do not receive live evaluation events yet; connection tests are available.
Model capabilities
The active backend determines support for reasoning settings, JSON schemas, tool parameters, and images. The gateway validates common request fields and forwards supported protocol features for backend validation. An accepted field can still be rejected for a particular model. The current image-capable model isabliterated-model; the large models are text-only. Use the live model catalog and verify your chosen combination. See images, reasoning, and web tools.
Protocol boundaries
- Responses are stateless:
store: true,background: true,previous_response_id, and stored-response retrieval/deletion are unavailable. Send conversation history directly. - Inference bodies are JSON with an 8 MiB maximum. Output bounds follow the configured model maximum. Chat
naccepts 1–128; other surfaces return one response. - Supported hosted tools are web search on Chat/Responses/Messages and web fetch on Messages. Custom function execution remains your application’s responsibility.
- Embeddings, audio generation/transcription, image generation, files, batches, fine-tuning, Assistants, and Realtime are outside this API surface.
- Organization APIs act on Limitless tenant resources, balances, and policies. They do not expose shared upstream account administration.
- Console sections still marked Preview remain previews; API coverage does not imply a completed management UI for each resource.
Streaming and accounting
The API and console Playground stream text. The Playground supports stopping generation while retaining partial output. A200 only confirms that a stream opened; inspect the terminal event and Activity for its outcome.
The gateway requests final Chat usage, records reported tokens from all three protocols, and applies cached-input rates. Unknown usage can retain a reservation for reconciliation. Hosted tools currently add no separate fee beyond reported model tokens.
Every response includes a gateway X-Request-ID. The gateway does not automatically retry inference, and inference POST requests have no idempotency guarantee. Administrative creation has separate idempotency requirements. See errors, billing, and administration.