Skip to main content
Use this guide when you need to answer one of three questions: is Phaseo healthy, is a model routable now, or does a route meet your application’s reliability target?

​
Check current service health

  1. Open the Phaseo status page for current incidents and incident history.
  2. Call GET /v1/health for a minimal Gateway health check.
  3. If one request failed, keep its request ID and inspect the activity or generation record before assuming the whole service is unavailable.
The public status page reports service health. Phaseo does not currently claim a contractual public uptime SLA.

​
Confirm that a model is routable

The Phaseo catalog is broader than the set of models available through the Gateway. Catalog presence means Phaseo tracks the model; it does not mean a public route exists. GET /v1/models returns only currently routable models by default:
Use availability=all only when you intentionally want to inspect coming-soon or inactive mappings. Do not send production traffic to a provider row unless availability_status is active and is_active_gateway is true. Availability does not guarantee perpetual uptime, access in every region, or identical commercial terms for every workspace.

​
Interpret performance metrics

Phaseo performance data is observational: it summarizes requests routed through the Gateway rather than a controlled lab benchmark. TTFT and output speed require a streaming response with a content-bearing first output. A non-streaming request can still contribute duration and effective-throughput measurements without an invented TTFT. Always read a metric with its time window, route, region, streaming mode, and percentile. Provider load, prompt length, output length, retries, and transport conditions can all change the result. For the complete definitions, see Pricing & Performance and How Phaseo measures latency and throughput.

​
Validate with your workload

Public telemetry is directional evidence, not proof that a route meets your production target. Build a small, reproducible study before a rollout:
  1. Remove sensitive production data and keep a representative prompt mix.
  2. Record the model ID, endpoint, provider constraints, region, streaming mode, and concurrency.
  3. Run enough repetitions to compare distributions rather than one request.
  4. Capture success rate, Gateway TTFT, provider duration, Gateway E2E, output tokens, and final cost.
  5. Test invalid keys, invalid models, rate limits, and unavailable-provider failures.
  6. Save request IDs and the exact time window so another engineer can reproduce the result.
Do not publish a customer name, quote, workload result, or reliability percentage without permission and a documented method.

​
Triage a failed request

  • Check service status.
  • Confirm the model still appears in the default GET /v1/models response.
  • Review the HTTP status and error code in Error Handling.
  • Retry transient 429 and 5xx responses with bounded exponential backoff.
  • Keep the request ID when contacting support.
Last modified on August 30, 2026