Routing policies

Route by cost, speed, quality or context — in a file you can commit.

A policy is an ordered list of rules. The first rule whose when matches decides the model; if none match, the default rule does.

Shape

name: production
rules:
  - when: task == "summarize"
    then: cheapest_above(quality: 80)
  - when: context_tokens > 400_000
    then: model("google/gemini-3.1-pro")
  - when: p95_latency > 800ms
    then: fallback_next
default: best_quality(max_input_price: 5.00)

Selectors

Selector	Picks
`cheapest_above(quality)`	The lowest blended price with a score at or above the floor
`fastest(within_quality)`	The lowest p50 first-token latency within N points of the best score
`best_quality(max_input_price)`	The highest score that costs no more than the cap
`longest_context`	The largest context window in the pool
`model(id)`	Exactly that model

Task detection

The task field is inferred from the request unless you set the x-prism-task header. Detection is a small classifier that runs in the router, adds under a millisecond, and is logged so you can check it.

Committing policies

Policies are YAML. Keep them in your repo and push with prism policies push; the dashboard editor reads and writes the same format.