What happens to my app when an AI provider goes down?
If you've added backup models to your workflow, ModelRiver retries the request on the next model automatically. Your app gets an answer from the backup instead of an error, and it doesn't need any retry code of its own.
When does a backup take over?#
A backup is tried when the current model:
- returns an error
- times out
- hits the provider's rate limit
- returns output that doesn't match your structured output schema
How many backups can I add?#
Up to two per workflow. They can be other models from the same provider or from a different one, for example Anthropic first, then OpenAI, then Google. Using another provider gives the best protection against a full outage.
Does a failover cost me extra requests?#
No. A request that fails over counts once toward your ModelRiver plan. Each provider that actually produced tokens bills you for them as usual.
How do I know a backup was used?#
Every response lists each attempt in meta.attempts. In Request Logs, the Timeline view shows which models were tried, why each failed and which one answered. It's a good place to spot a provider that's struggling.
What if every model fails?#
The request returns an error with details of each attempt, so your app can show a friendly message or try again later. A spend limit can also stop requests on purpose; that's a budget error, not an outage.