Methodology

Numbers without context are marketing. This page documents how the latency, pricing, retention, and SLA claims across Router One are actually measured — so you can verify or reproduce them, not just take our word.

Last updated: 2026-05-12

Latency tests

Sampling window: Rolling 7-day window. Numbers on landing pages are refreshed monthly; the dashboard updates continuously.
Cities: Beijing, Shanghai, and Shenzhen — chosen to span the three main domestic backbone networks.
ISPs: China Telecom, China Unicom, and China Mobile. Each city runs probes on all three operators.
Probe type: End-to-end HTTPS request to api.router.one with a representative chat completion payload (~500 input tokens, ~200 output tokens, non-streaming).
Sample size: ≥ 5,000 requests per city-ISP pair per week. Outliers above p99.5 are reported separately, not folded into the headline median.
Numbers we publish: Median (p50), 95th percentile (p95), timeout rate. Latency on landing pages is the median across all cities and ISPs unless the page is region-specific.

Pricing & rate updates

Per-model rates: Input, output, and cached-input prices on /models are the rates you are billed at. Any upstream price change is reflected within 24 hours of the upstream provider publishing it.
No hidden markup: Pay-as-you-go calls are billed at posted rates with no per-token markup. Subscription plans apply a flat discount that is shown on /pricing.
Currency: USD-denominated internally; conversion to RMB for WeChat Pay and Alipay top-ups uses the daily mid-market rate plus the channel's fee, both shown at checkout.
Notification of changes: Material rate changes are announced via the changelog and email to active accounts at least 7 days in advance.

Data retention

Request and response bodies: Not stored. Router One proxies requests and responses in real time without persistent capture of message content.
Metadata logged: Per request we record: timestamp, project and API key ID, model used (and fallback chain), input/output/cache token counts, latency, status code, and cost. This is what powers the trace and billing.
Retention window: Metadata is retained for 90 days for observability and 13 months for billing reconciliation, then purged. Enterprise contracts can extend or shorten this.
Upstream behavior: Router One does not control what the upstream provider does with the request. We pass it through; their data policies apply on their end.

Availability & SLA

Reported uptime: Monthly availability based on synthetic probes (one HTTPS request per minute from each city). Numbers exclude scheduled maintenance announced ≥ 48 hours in advance.
Failover behavior: When an upstream provider's error rate or p95 latency crosses configured thresholds, smart routing falls back to an alternate provider. The fallback is recorded in the per-request trace.
Status page: Live incident status is published on the status page (link in the footer) and to subscribed email addresses.
Enterprise SLA: Annual contracts can include a written SLA with credits for missed availability. Contact sales via /contact.

Questions about the methodology? Email us via /contact.