Claude Code cost per task: real proof
The most reasonable question before using Claude Code for work: “one task done, how many rupiah?” Few people answer it; most just say “it depends”. True, but useless.
This guide gives you three things: the cost formula, an illustrative worked example from the openly displayed catalog prices, and our honest status on the real measurement we are preparing.
The status of the numbers in this article: read first
We are preparing one real measurement: a coding task run in Claude Code through Giant Router, with starting balance, ending balance, and a usage report screenshot. The number will be published in the Cline demo guide once the measured run is finished. We do not want to publish numbers that have not been measured.
So in this article, every number below is a worked example (illustration) from catalog reference prices, not a measurement. The arithmetic is exactly what you will see in your own usage report.
The cost formula
Every request is charged by token:
cost = (input tokens / 1M x input price) + (output tokens / 1M x output price)For agents like Claude Code, one “task” is not one request. A task usually consists of several requests (read files, think, edit, run tests), and each request re-sends the conversation context as input tokens. That is why agentic tasks cost more than a plain chat of the same length: input tokens dominate.
Reference prices and a worked example (illustration)
From Giant Router’s reference price catalog (public demo catalog snapshot 2026-11-08; prices can change, rupiah conversion uses the display rate Rp16,500/USD):
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| claude-haiku-5-5 | $0.10 | $0.50 |
| claude-sonnet-5 | $2.00 | $10.00 |
Worked example: an illustration, not a real bill.
Small task: fix one function plus one test. Estimated tokens: 40,000 input, 2,000 output.
- claude-haiku-5-5: (40,000/1M x $0.10) + (2,000/1M x $0.50) = $0.004 + $0.001 = $0.005 ≈ Rp83
- claude-sonnet-5: (40,000/1M x $2.00) + (2,000/1M x $10.00) = $0.08 + $0.02 = $0.10 ≈ Rp1,650
Medium task: refactor one module with several test iterations. Estimated tokens: 200,000 input, 8,000 output.
- claude-haiku-5-5: $0.02 + $0.004 = $0.024 ≈ Rp396
- claude-sonnet-5: $0.40 + $0.08 = $0.48 ≈ Rp7,920
An honest note about this example: the token counts above are shape-of-usage estimates, not measurements. The shape is what matters; you can see the “input dominates” pattern and why a 20x price difference between models moves a task’s cost from hundreds to thousands of rupiah. The exact numbers come out of your usage report and (soon) our demo guide.
Why you can rely on this at Giant Router
Two structural reasons, not promises:
1. Per-key budget limits. You set a limit (say Rp50,000) on the key Claude Code uses. Usage stops at that limit. Even the most wasteful task cannot pull your balance past it; the balance cannot go negative.
2. Per-request reporting. Every request is recorded: model, tokens, cost. After a week of work you can total up “this month’s coding cost” yourself instead of guessing. If you bill a client, this same report is your reimbursement attachment; how to use it is covered in billing clients for AI API usage.
How to compute your own task cost
- Run one task in Claude Code with a Giant Router key.
- Open the usage report, note the task’s total input and output tokens.
- Plug them into the formula above using your chosen model’s price (reference prices are on the homepage).
- Compare with your earlier estimate. The gap is usually in the number of requests, not in the per-token price.
After a few tasks you will have a personal number that is far more useful than any article: the task cost of your own codebase.
Two questions that come up often
“Isn’t a monthly subscription cheaper than paying per token?” It depends on your usage, that’s the point of computing. “Unlimited class-X Claude” monthly packages on the market generally sell in the Rp49-50k/month range (a market price we observed for our own pricing decisions). If the worked example above puts your monthly tasks below that, pay-per-token is cheaper, and if your usage is far above it, the reverse. The difference from a subscription: here the number comes out of your own report, not from a brochure’s assumptions.
“Why can the same task cost differently each time it runs?” Reasoning models produce “thinking” tokens that are also billed, and their volume varies per request. So an identical task run twice can cost different amounts, not a reporting bug, just the nature of generative models. Per-request reports show this pattern, and if consistency matters more than maximum quality, a cheaper non-reasoning model is usually the sensible choice.
Summary
The formula is simple, the worked example is repeatable, and two safety nets (per-key budget limits and per-request reporting) make “what did this task cost” a question that always has a written answer. We are preparing the real measured number and will publish it in the demo guide; until then, run one task yourself and watch the number appear.