Routing policies

Cheap-good-enough by default, frontier only when the caller says it matters. The 40% of your bill that was reflex, deleted.

{
  "name": "frugal-coder",
  "rules": [
    { "name": "explicit-priority", "match": { "metadata": { "priority": "high" } },
      "route": { "model": "claude-fable-5" } },
    { "name": "long-context", "match": { "prompt_over_chars": 8000 },
      "route": { "model": "claude-sonnet-4-6" } }
  ],
  "default": { "model": "claude-haiku-4-5-20251001" }
}

Start cheap; if the session is still going after four turns, the task earned a better model. Pay for difficulty, not by reflex.

{
  "name": "escalate-on-struggle",
  "rules": [
    { "name": "earned-the-frontier", "match": { "turns_over": 4 },
      "route": { "model": "claude-fable-5" } },
    { "name": "warming-up", "match": { "turns_over": 1 },
      "route": { "model": "claude-sonnet-4-6" } }
  ],
  "default": { "model": "claude-haiku-4-5-20251001" }
}

Claude does the work; anything tagged for review gets re-run through the OpenAI harness. Two vendors, one transcript — the referral pattern as policy.

{
  "name": "second-opinion",
  "rules": [
    { "name": "review-pass", "match": { "metadata": { "stage": "review" } },
      "route": { "harness": "codex", "model": "gpt-5" } }
  ],
  "default": { "harness": "claude", "model": "claude-sonnet-4-6" }
}

Customer-facing chat rides the budget tier; coding sessions get the mid tier. The classic platform split, two rules.

{
  "name": "chat-cheap-work-mid",
  "rules": [
    { "name": "chat", "match": { "workload": "chat" },
      "route": { "model": "claude-haiku-4-5-20251001" } },
    { "name": "coding", "match": { "workload": "coding" },
      "route": { "model": "claude-sonnet-4-6" } }
  ],
  "default": { "model": "claude-haiku-4-5-20251001" }
}

Your paying tenants get the frontier; free-tier users get good-enough. Per-user tiering with the metadata your app already sends.

{
  "name": "vip-lane",
  "rules": [
    { "name": "paid-tenant", "match": { "metadata": { "plan": "pro" } },
      "route": { "model": "claude-fable-5" } }
  ],
  "default": { "model": "claude-haiku-4-5-20251001" }
}

One line: let cerver's automatic router pick per turn — coding signals get the strong path, research goes to a current-answers route, routine work rides the cheap tier. Your rules (if any) still win first; auto only fills what's left. Opt-in by design: no policy means nothing ever changes behind your back.

{
  "name": "cerver-auto",
  "auto": true,
  "rules": []
}