What Claude Code bills
Every turn, Claude Code sends the whole conversation back to the model: the system prompt, tool definitions, the files it has read and everything said so far. On the API you pay for four kinds of tokens.
| Claude Opus 5.5, per 1M tokens | Anthropic list | Zurelay |
|---|---|---|
| Input (new) | $4.00 | $1.40 |
| Cache write (5 minutes) | $5.00 | $1.75 |
| Cache read | $0.40 | $0.14 |
| Output, thinking included | $20.00 | $7.00 |
Cache reads are most of the volume: the start of each request repeats the last one, so it comes back from the cache at a tenth of the input price. Output is the most expensive per token, and thinking at high effort counts as output.
A worked example
Take a busy month of agentic coding: 5,000 requests of about 40,000 tokens in, 85% of them read from the cache, and 1,500 tokens out. Your numbers will differ, so treat this as a shape, not a quote.
| Model | At Anthropic's list price | On Zurelay |
|---|---|---|
| Claude Opus 5.5 | $338.00 | $118.30 |
| Claude Sonnet 5.5 | $169.00 | $76.05 |
| Claude Haiku 5.5 | $8.45 | $3.38 |
This counts new input at the input price. Claude Code writes much of it to the cache at 1.25 times that, so a real bill runs a little higher. The cost calculator takes your own numbers.
How to spend less
- Map the Haiku alias to a cheap model. Claude Code uses it for background work. Claude Haiku 5.5 costs $0.20 per 1M output tokens.
- Use Sonnet for routine work and keep Opus for hard problems: set the model per session with
--model. - Keep sessions short. Every turn re-sends the conversation, so start a new session for a new task and compact long ones.
- Lower the thinking level for simple edits: thinking is billed as output.
- Pay less per token. On Zurelay every Claude model is below Anthropic's list price (65% on Opus 5.5), through the same Messages API, so Claude Code works after setting five environment variables. See Claude Code for less.
Subscription or API?
Claude's subscription plans include Claude Code with usage limits that reset on a schedule. The API has no limit beyond your balance and bills each token. The API suits teams sharing one balance, Claude Code in CI, and anyone who keeps hitting plan limits.