People ask how we can charge less than list price for the same models. Here's the honest, mechanical answer — no magic.
Every model has many routes
A model like Claude Opus 5 isn't one thing. It's served through multiple upstream routes — different deployments with different availability, latency and cost. When you call a model through NewToken, there's a decision to make: which route serves this request?
We expose that decision. In the catalog, every model lists its available routes with per-route pricing and live online/offline status. By default your requests take the cheapest online route — that's the "auto" you see in the key editor.
Health-aware, not price-blind
Cheapest is only useful if it works. Routes go unhealthy — they rate-limit, they slow down, they go dark. Route status is checked continuously, and auto-routing only considers routes that are currently online.
If you'd rather pin a specific route (say you benchmarked one as faster for your workload), the key editor lets you pin any route per model, per key. Pinning is yours; auto is ours.
Where the discount comes from
We aggregate demand across all our users and route it efficiently. Upstream economics reward volume and flexibility; we pass most of that back as the discount you see on every model card. Our cut is the margin between our route cost and what we charge you — it's visible (list price crossed out, our price, the percentage), and it's how the platform funds itself. No platform fee, no seat pricing, no hidden markup in the token math.
What this means for you
- The price on the model card is the price you pay. It already includes everything.
- If a route degrades, your traffic moves (auto) or you move it (pinned).
- Your bill is metered by the upstream's token counts — the same metering the vendor would use on you directly, just cheaper.
Questions about a specific model's routes? Open its page in the catalog — everything we know about its routes is right there.