The provider most likely to solve.

Ranks providers by success rate over recent attempts, discounted for small samples.

Routing

Worked example, computed by the routing function: a perfect record of 10 ranks below 980 of 1,000.
Windows
48h·30d
Attempts a window needs
10
Ranked on
Lower bound
Volume weighted
Per organization

01Rule

Consistency beats a lucky streak.

  1. 01Count every attemptA failure is scored against the provider that ran it, even when a retry or fallback later solves the task.
  2. 02Weight by organizationEach organization counts by the square root of its volume, so one broken setup cannot sink a provider for everyone.
  3. 03Rank on the lower boundThe Wilson lower bound of the success rate, freshest qualifying window first.
  4. 04No window qualifiesPrice order, as Auto Cheapest.

02Config

One field turns it on.

settings.routing, in POST /createTask

{
  "settings": {
    "routing": {
      "mode": "autoMostReliable"
    }
  }
}
How Auto Most Reliable ranks
SettingValue
ModeautoMostReliable
MetricWilson lower bound of the success rate
RecomputedEvery 15 minutes
Without dataPrice order

03FAQ

Questions about Auto Most Reliable.

01

How is success rate calculated?

Solved attempts divided by all attempts in the window. Providers are ranked on the lower bound of that rate, so a small lucky sample does not outrank a proven one.
02

What counts as a failure?

Every attempt that does not return a solution, retries and fallbacks included. Each is scored against the provider that ran it.