# Router One Smart Routing & Fallback Facts

Last updated: 2026-07-13

- Product: Router One
- Category: LLM API Gateway / AI Model Router
- Default routing strategy: model_name; requests that specify an exact model do not automatically enter adaptive candidate scoring
- Adaptive model selection: model="auto" uses a server-owned candidate set and the auto_route retry budget; it does not accept a customer-defined router object
- Adaptive routing signals: latency, cost, reliability when a supported candidate-routing path is used
- Latency signal: EWMA keyed by model_provider, with alpha=0.20 and a 15-minute metric TTL; it is not a fixed last-50-request window
- Candidate handling: model="auto" can retry another server candidate; an exact-model request can retry another healthy provider route serving that same model
- Exact-model fallback scope: provider retries keep the requested model unchanged; the public contract does not promise cross-model substitution for that path
- Unsupported controls: project_weights, api_key_weights, request_router_field, project_fallback_override
- Unsupported guarantees: last_50_window, quality_score, fallback_within_200ms
- Latency boundary: fallback includes the failed attempt plus any retry; there is no universal fixed fallback latency guarantee
- Trace contents: the final model and provider route, input/output tokens, cost, latency, and status for the request; intermediate failed-attempt chains are not part of the public trace contract
- Trust boundary: smart routing selects upstream routes for LLM API requests; Router One is not an agent orchestration runtime or workflow engine

## Canonical routing pages

- Smart routing methodology: https://router.one/routing-methodology
- Automatic LLM fallback: https://router.one/llm-fallback
- LLM API gateway overview: https://router.one/llm-api-gateway (markdown: https://router.one/llm-api-gateway.md)
- China latency benchmark: https://router.one/benchmarks/china-latency
- Company facts: https://router.one/facts/company.md