Skip to content

Fallback

Providers come and go. When none can take your request, Katara can answer from a managed model instead of failing. Today that model is DeepSeek.

Only after the marketplace is exhausted for a request: no live offer for the model, or every attempt failed before the first token. The fallback never competes on price with a provider that can serve you.

The reply’s system_fingerprint starts with fallback: and the katara object has fallback: true. In the chat app the cost line says “fallback”.

Each model has a published fallback price. It is charged like any other reply, from your wallet, with the same fee. The admission check applies too: with no spendable balance there is no fallback.

If a provider fails after it has already sent part of a reply, the reply ends with an error and nothing is charged. Katara does not switch models mid-answer. Send the message again; by then the failing provider is on hold and the fallback takes over.

Models without a published fallback report no capacity when no provider is live.