Where Does a Claude Code Bill Actually Go?

Every line is three numbers

A coding-agent request is billed on three components. Build the reflex: when a bill surprises you, don't ask which model is expensive — pull up lines and ask which of the three numbers moved.

1. Prompt tokens — everything the model reads: system prompt, project context, the conversation so far, tool results. On a long session this is almost never "your question"; it's your question plus everything the agent re-read to answer it.
2. Completion tokens — what the model writes back: code, reasoning, tool calls. Usually the minority of an agent bill.
3. Cache reads/writes — prompt-cache pricing. Reads are steeply discounted hits; writes populate the cache. Whether your tool and endpoint cache well moves your effective rate dramatically.

The quiet dominant: context gets re-sent

Claude Code and similar agents are stateless at the wire: each turn, the client sends the conversation again — system prompt, file context, prior tool results included. Your 15th message in a session is billed as if you asked all 15 at once. That's not a flaw; it's how the protocol works. But it explains two counterintuitive facts:

A long cheap-model session can outcost a short expensive-model one. And the fix with the best ROI is usually managing context — clearing stale tool output, keeping sessions scoped, letting cache do its job — not hunting discounts.

Three levers, in order of honesty

Before touching a rate table, these are the levers ranked by how honestly they save money:

1. Right-size the model per task. Boilerplate edits and greps don't need the flagship; the hard architectural asks do. Routing by task is the biggest honest saving.
2. Make caching actually work. Prompt-cache reads are an order of magnitude cheaper than cold prompt tokens. Stable system prompts and session structure matter more than any rate table.
3. Subscription vs. usage — do the arithmetic for your shape. Plans win if your usage is bursty-heavy on one model; usage-based wins the moment your workload is diverse, occasional, or shared across tools. There is no universal answer, only your own math.

Three questions to ask any provider

Whatever endpoint you use, these turn a black box into an audit trail:

1. Can I see per-request cost, tokens, and latency — not just a monthly total? You should be able to export it.
2. Is there a machine-readable price list, so scripts can compute cost before committing? A public no-login JSON endpoint is the bar I'd hold any provider to.
3. What exactly happens to unused credit — expiry, refund policy, fees? "It just sits there forever" should be a written policy, not a vibe.

Privacy, since you're routing code through an endpoint

Ask what's retained. Teapot Router keeps billing metadata only — no prompt/response archiving — and you should verify rather than trust any vendor's "zero log" marketing, including ours.