The same model is wired to more than one upstream route
Routes track health independently, so a problem on one doesn't affect the others.
When an upstream provider goes down,
your users shouldn't have to know.
The same model is wired to more than one upstream route. When one fails, the same request automatically switches to the next — not starting over, picking up where it left off.
How routing picks, for a single request
The primary and fallback routes track health separately; whether to switch is decided by that request's own health check, not a bulk retry policy.
Long-running jobs from real customers
Text streams as long as 60 minutes, video jobs as long as 9.8 hours — both ran to completion.
60 minutes
Longest observed text stream
67,462 tokens, completed successfully
9.8 hours
Longest observed video job
Asynchronous generation, completed successfully
Historical platform data, not a service commitment.
For long-running jobs, we recommend api.bazaarlink.ai — it's a dedicated API endpoint deployed separately from the website, so it isn't interrupted by website deploy restarts.
How it works
Four facts about the mechanism, each mapped to an actual request path — not marketing copy.
Circuit breaking only trips the failing route
One route having problems won't take the whole model down, and other models are completely unaffected.
Retries happen before you see any content
We confirm a model is actually responding before we start streaming content to you — once you've started receiving an answer, the provider won't switch mid-stream.
Failures show up in your usage log
You can see that a retry happened and see the failure code — it isn't swallowed silently; the log won't reveal which upstream provider was actually used.
If the connection drops after content has already started
In this case, we won't switch you to another provider and retry. What you'll see is a clear error, not a response that looks complete but actually broke off midway — and we won't charge you for that request either. This describes a mechanism, not an availability guarantee — if your contract includes an SLA, the contract governs.
Want the full routing and circuit-breaking logic?
How routes are ordered, how the circuit-breaking threshold is calculated, and the field definitions in the usage log — it's all in the docs.
Point your own traffic at it
Five minutes to hook up — a single route failing won't be something your users notice.
See how failover works