• intelligent routing •
The router that picks
the model for you.
Send model: routerbench/auto and RouterBench classifies each request, then routes to the best model on quality, cost, latency and reliability — and fails over automatically when a provider degrades.
• how it decides •
Every request, classified and scored.
01 — classify
12 task types, 4 domains
Chat, code, SQL, math, reasoning, extraction, summarization, translation, classification, tool use, creative — plus finance, healthcare, legal and science sensitivity.
03 — profiles
Pick a priority
Cost saving, balanced, best quality or high reliability — one setting, the router does the rest.
05 — flags
Per-request steering
free_models_only, no_reasoning, or reasoning_effort — take the wheel when you want it.
06 — on-perimeter
Classification stays in-perimeter
Query understanding runs inside your own network — prompts never leave to a third party to decide where they go.
provider scoring weights
• provider selection •
Then it picks the best host, too.
- scored — When a model has multiple hosts, each provider is scored — best one wins
- failover — Below 90% uptime triggers automatic failover to the next-best host
- resilient — Retries with exponential backoff and circuit breakers absorb failures
- fresh — ~1% exploration traffic keeps performance data current
- governed — Managed globally, or org-controlled per project
Route on data, not gut feel.
Point at routerbench/auto and let the router earn its keep on your traffic.