Routing & failover
You call one canonical model id through one stable endpoint. AIx evaluates the eligible paths for that request and selects a healthy route automatically. If the first path fails, AIx transparently tries the next one.
The AIx route engine
Route selection is capability-aware. Every request is evaluated against:
- Model availability — the requested model must be active on the route.
- Feature support — tools, structured output, vision and streaming require compatible paths.
- Live health — degraded routes and open circuits are skipped.
- Efficiency — cost and latency guide selection between healthy candidates.
The catalog and route health are managed by the AIx control plane. New model availability can be activated without changing your integration or redeploying your application.
Automatic failover
For each request, AIx walks the eligible route set:
- Skips any route whose circuit is open.
- Sends the request through the best healthy candidate.
- Records unsuccessful attempts and tries the next eligible path.
- Returns
503 service_unavailableonly when no healthy path can serve the request.
This applies across supported chat, embeddings and media workloads. Your request and response shapes remain stable throughout the process.
Circuit breaker
Failover alone would keep sending traffic to an unhealthy route. AIx uses a distributed circuit breaker so all server instances share the same health state.
| Setting | Value |
|---|---|
| Failure threshold | 5 failures |
| Counting window | 60 seconds |
| Open cooldown | 30 seconds |
When a route crosses the threshold, its circuit opens and AIx temporarily removes it from selection. After the cooldown it is safely probed again; a successful response clears the failure counter.
Advanced request routing
Requests using tools, structured outputs (response_format) or multimodal content
are restricted to routes that support every requested capability. See the
Compatibility layer for the feature model.
Choose the model for the capability and quality you need. AIx owns route selection, health checks and failover so your application does not need infrastructure-specific logic.