Skip to main content
orcarouter/auto is a named router we create for every account on signup. It routes each request to the cheapest live chat model your account has access to, chosen fresh per request.

​
Usage

response = client.chat.completions.create(
    model="orcarouter/auto",
    messages=[{"role": "user", "content": "..."}],
)
No other setup required — the router exists the moment your account is created.

​
Default behavior

The seed configuration:
  • Pattern: empty — matches every chat model your account has access to. New models that come online become candidates automatically.
  • Strategy: cheapest — picks the model with the lowest per-token price among live candidates.
  • Default model: none. If the pattern resolves to no available models, the request fails with a clear error. You can set a default in the dashboard if you want a safety net.
You can see and edit your Auto Router in the dashboard under Routing. You can narrow the pattern (e.g. restrict to openai/*), swap the strategy, set a default_model, or delete the router entirely — same as any named router.

​
Switching strategies

The dashboard exposes four strategy cards. Cheapest is the seed default above; pick another to change how orcarouter/auto resolves picks:
CardBackend enumWhat it does
CheapestcheapestLowest per-token price among live candidates. The seed default.
QualityqualityHighest quality score among live candidates, regardless of price.
BalancedbalancedPicks a low-cost option that still meets a quality bar; falls back to the highest-quality option if nothing meets the bar. Default for new routers you create yourself.
Adaptivelinucb / gated_adaptivePer-router LinUCB contextual bandit that learns from your real traffic to weigh quality, cost, latency, and reliability per request. Two sub-modes (Standard / Gated) — see Named Routers.
Adaptive needs a short warm-up per model before it starts steering picks. During warm-up it behaves like Balanced — that’s expected, not a bug.

​
When to prefer Auto Router over explicit model names

  • You don’t want to pin to a specific model; you want the cheapest live chat model at each request.
  • You’re prototyping and don’t want to care about which provider is up.
  • You want OrcaRouter’s routing to “just work” without thinking about it.

​
When to prefer explicit model names

  • You need deterministic output — picking different models at different times will change generation style and quality.
  • You’re using features specific to one model (e.g. Claude’s cache_control, or a model’s native image generation).
  • You want predictable per-request cost.

​
Seeing what Auto Router picked

Check the X-Orca-Resolved-Model response header. See Response Headers. 
res = client.chat.completions.with_raw_response.create(
    model="orcarouter/auto", ...
)
actual_model = res.headers.get("X-Orca-Resolved-Model")
# e.g. "openai/gpt-4o-mini"