The single biggest line item in most AI bills is habit: wiring every call to the flagship model because it's the one you tested with. Routing each request to the cheapest model that can actually do the job is usually a one-word change — and it compounds fast.
Key takeaways
- Not every task needs your most expensive model.
- Route cheap, high-volume work to small models; save flagships for hard jobs.
- On a gateway, switching models is one string per request.
- The savings scale with volume — this is where real money is.
The overpaying trap
You prototype on the best model, it works, and the model name never changes. But classification, short replies, extraction and routing don't need frontier reasoning — a mid-tier or small model does them just as well for a fraction of the price.
A simple routing policy
| Task | Model tier | Why |
|---|---|---|
| Hard reasoning, long agents | Flagship (Astra / Opus 5) | Worth the price on genuinely hard work |
| General everyday work | Workhorse (GPT-5.6 / Sonnet 5) | Strong at a fraction of flagship cost |
| Classify, extract, short replies | Small (mini / nano / Haiku / Fable) | Fast and cheap; the job is easy |
Start by moving your highest-volume endpoint down one tier and measuring quality. If it holds, you just cut that path's cost by a large multiple.
Why a gateway makes this trivial
Routing is only easy if switching models is easy. Through an AI gateway, every model is the same OpenAI-shaped call — so your routing logic is just choosing a string:
const model = isHard(task) ? "claude-opus-5" : "claude-haiku-4-5";
await ai.chat.completions.create({ model, messages });Same key, same bill, with a spend cap underneath so a mistake can't run away.
What it adds up to
Dropping a high-volume path from a flagship to a small model can cut that path's token cost by 10× or more — with no user-visible change.
Compare the real per-token prices across tiers on the models page, pick where each task belongs, and let the gateway do the switching.
The cheapest optimization is sending cheap work to cheap models. See the prices and start routing.