System Status & SLA

zrouter exposes live operational data publicly. This page explains what is published, how the gateway behaves when the upstream provider has trouble, and our commitment around service levels.

The status page shows live uptime over the trailing 24 hours, 30 days, and 90 days, plus per-model time-to-first-token (TTFT) at the p50 and p95 percentiles. It is updated continuously and reflects real traffic, not synthetic probes.

Internally, zrouter exposes Prometheus metrics — request counts, latency, upstream first-byte time, and server overhead — from a dedicated metrics endpoint. The gateway itself runs as redundant replicas behind a load balancer, so the loss of a single replica does not take the API down.

When the upstream provider becomes flaky, the gateway degrades deliberately rather than failing loudly or billing you for nothing:

• Automatic retry before the first byte. Transient upstream failures are retried a small number of times before any output is sent to you.
• Circuit breaker. If the upstream is down for long enough, requests fail fast with 503 circuit_open and a Retry-After header instead of hanging.
• No mid-stream retries. Once streaming has started, the gateway never retries — a stream interruption is returned to you as-is, and because settlement is debit-after, a truncated stream is not over-billed.

We are transparent about real availability: we publish live uptime and latency on the status page and disclose incidents as they happen. We do not publish contractual uptime targets today; formal SLA terms will be published when they become available.