The Multi-Provider Router

• I run four practices. My business model is that I refuse to specialise.· A lawyer who builds software, a producer who reads Friston, a writer with two bar memberships. The orthodoxy says pick one. The numbers say something else.
• The Reel Factory· The protocol that built the reels you're watching.
• The Bias Mirror· Use AI as critic, not cheerleader. Friston's price for honesty.

typescript
// Router classifies the task profile, then routes.
async function route(task: Task): Promise<Response> {
  const profile = await classify(task);

  if (profile.simple_extraction || profile.classification) {
    return call("haiku-4.5", task);            // $0.80/M
  }

  if (profile.huge_context && !profile.reasoning) {
    return call("gemini-2.5-flash", task);     // 1M ctx, $0.30/M
  }

  if (profile.multi_step_reasoning) {
    return call("sonnet-4.6", task);           // $3/M
  }

  if (profile.frontier_coordination) {
    return call("opus-4.7", task);             // $15/M
  }

  // Fallback chain: try cheap first, escalate on confidence drop
  return tryChain(["haiku-4.5", "sonnet-4.6", "opus-4.7"], task);
}

• embedding-based classification
• prompt-token vs completion-token economics
• provider SLA differences (rate limits, region pinning)
• the latency-cost-quality trilemma
• fallback chain design
• LiteLLM / OpenRouter proxy patterns
• the 80%-of-traffic-on-cheap-tier rule

If a name is unfamiliar, that's the gap. The list is the curriculum.

1. 01

   Audit your last 1000 requests. Tag each as classification, extraction, reasoning, or coordination.

2. 02

   Train (or fine-tune a prompt for) a classifier that returns the profile in ~30 tokens.

3. 03

   Build the routing table. Cheapest model capable of each profile wins by default.

4. 04

   Add a confidence check. If the cheap model returns low confidence, escalate one tier. Never two.

5. 05

   Pin a quality canary. 1% of traffic always goes to the premium tier; compare outputs daily.

6. 06

   Track cost-per-task and resolution-rate per route. The trilemma is real; the dashboard makes it negotiable.

7. 07

   Re-tune monthly. New model releases invalidate the table.

Most teams pick one model and pay frontier prices on extraction tasks. Routing flips the default: cheap unless proven otherwise. The classifier costs ~30 tokens; the savings on a single rerouted task usually exceed 100 tokens. At any volume the maths only goes one way. The discipline is the canary, not the routing; without it you can't tell when a tier change degrades quality silently.

• Anthropic API pricing
  Anthropic · primary source
  Live pricing for Opus / Sonnet / Haiku tiers.
• Gemini API pricing
  Google · primary source
  Live pricing for Gemini 2.5 Pro / Flash / Flash-Lite.
• LiteLLM proxy
  BerriAI · primary source
  Provider-agnostic routing layer. Makes the cheap-by-default discipline easy to ship.