How routing works
The complexity router, prompt caching, and multi-host sourcing that lower your bill — with worked arithmetic and open methodology, not hype numbers.
We show the arithmetic, not a marketing number.
Our margin comes from buying tokens smarter than you would, not from a markup — that is what funds at-list pricing with no top-up fee. Here is exactly how it works, and how we measure it. No “500× cheaper” claims; every number below is arithmetic you can check.
Which tier each kind of task needs
You always pay the published rate for whichever model answers. The saving is that most coding work does not need the frontier tier — so the question is which work does, and that is a policy you can read rather than a black box.
- Edits and renames Small diffs, narrow context, no planning. Routed to the air tier.
- Tests and fixtures Mechanical generation against a known shape. Routed to the air tier.
- Explain or review Reading, not reasoning across files. Routed to the air tier.
- Multi-file refactor Wide edit surface, cross-file consistency. Routed to the flagship tier.
- Architecture and planning Long-horizon reasoning; a cheap miss is expensive. Routed to the flagship tier.
Escalation is automatic: if a cheap tier's answer fails the task's check, the request is retried one tier up. Shares of traffic per class are not published here yet — they will be, from measured routing logs, rather than estimated.
UNVERIFIED The task-class to tier mapping describes intended router policy; verify against the live scorer before treating this as documentation.
Three mechanisms
Complexity routing
Each request is scored. Simple turns go to a model roughly 5× cheaper; hard turns stay on the full model. If a cheap answer looks incomplete, it escalates automatically.
Prompt caching
Coding sessions reuse a large, stable prefix. Cached input is billed at a fraction of fresh input, so a warm session costs far less than a cold one.
Multi-host sourcing
The same weights sell for different prices across hosts. We source from the cheapest host that passes our quality bench, and fail over if one degrades.
A worked example
Blended cost per million tokens for stratacode/glm-5.2 at a 4:1 input:output mix — that is 0.8 × input + 0.2 × output. Recompute it from the rates on the pricing page.
- List sourcing, cold, no routing$2.00—
- + 50% prompt-cache hit$1.54−23%
- + 40% of turns routed to GLM-4.5-Air$1.05−47%
- + sourced from a discount host that passed the bench~$0.63−69%
You are billed at list; the gap between list and our actual cost is what the stack saves. When a downgrade happens you are billed at the cheaper model’s rate, so part of that saving lands directly on your invoice.
The quality gate
Cheaper sourcing is worthless if the model gets quietly worse. Before any host or lane goes live it must pass a bench measuring tool-call fidelity, malformed-argument rate, latency and output-cap behaviour against a baseline. A host that regresses does not enter rotation. That is how we source cheaply without the quantized-junk problem that comes with blind aggregation.
What we report
Your dashboard shows real per-request numbers pulled from the metering ledger: tokens in, out and cached, which model answered, and the router’s counterfactual saving. Site-wide savings figures are not published here, because measuring them across all traffic has not started yet — when it has, the numbers will come from that ledger rather than from an estimate.