$2 a Day for Claude Code: A Budget, Not a Promise

01 / Start with a $2 limit, not an all-day promise

Leaving Claude Code available all day is different from keeping it working continuously. An idle terminal is not the same as a stream of model requests. Your bill depends on the requests made, their token counts and the rates charged—not simply how long the window stays open.

For a tiny-budget beginner, use $2 as a daily planning ceiling. Across 20 working days, that allows $40. That is budget arithmetic, not a claim that $2 buys eight hours of coding. A large repository investigation could consume your allowance much faster than several small edits.

The practical goal is simple: keep the assistant handy all day, but purchase narrowly scoped work in small, observable batches.

02 / Connect CheapRelay without guessing the endpoint

CheapRelay offers pay-as-you-go access with no monthly fee or subscription. Its API base is https://api.cheapapirelay.com, with OpenAI-compatible routes under /v1 and Gemini-native routes under /v1beta. Those route details do not mean every client uses the same configuration.

Sign up for CheapRelay, create an API key and follow its current Claude Code setup instructions. Use the endpoint and authentication settings specified for that client rather than assuming an OpenAI-compatible URL belongs in every configuration field. Keep your key out of source control and shared terminal screenshots.

Before starting real work, confirm the selected model and its current rates. Claude Haiku 4.5 is one option to evaluate for small tasks, but judge suitability with your own trial. Do not assume that a model name guarantees a particular price or that every task needs the same model.

03 / Spend the first $0.25 measuring one small task

Make your first session a measurement exercise. Set aside $0.25 of your $2 daily budget for a tightly bounded task: explain one function, identify the file responsible for a display bug, or propose one small refactor. This is a spending checkpoint, not a guaranteed task price.

Record your account’s reported spend before and after the request. If reporting is delayed, wait for it to settle before drawing conclusions. Also note the model, task size and whether the assistant inspected many files or repeated itself.

• ■ Ask for an explanation before authorizing a broad edit.
• ■ Name the relevant file or directory when you know it.
• ■ Request a short answer and a minimal patch.
• ■ Review the result before allowing another iteration.

If the trial costs more than expected, stop and narrow the next request. Do not extrapolate one cheap explanation into a promise about a repository-wide migration.

04 / Read the bill as four token buckets

CheapRelay bills real tokens, including cache reads and writes. Your useful cost estimate therefore needs more than the visible answer length. Track ordinary input, output, cache writes and cache reads using the categories and units shown in the current pricing information.

For rates quoted per million tokens, each category costs its token count divided by 1,000,000, multiplied by its applicable rate. Add the categories together. Avoid counting cached input twice; follow the provider’s definitions for its usage fields.

Caching may change the economics of repeated context, but cached does not mean free. Likewise, a short final answer can follow substantial input processing. The receipt is a better guide than the answer’s length.

05 / Split $2 into checkpoints, not unlimited loops

A workable daily allocation is $0.25 for calibration, $1.25 for planned tasks and $0.50 held back for corrections. These are personal envelopes, not automatic service limits. If spending controls are available, configure them; verify whether they actually stop requests or merely send alerts.

Otherwise, check usage between tasks and stop early enough to allow for delayed reporting or requests already underway. Avoid unattended loops on a tiny budget. Repeatedly asking the assistant to continue until everything is fixed gives you little control over the next bill.

06 / Keep the terminal open and the task list small

Work in a simple cycle: describe one change, inspect the proposed patch, run the relevant local checks and decide whether another paid request is necessary. Start a fresh session when the existing conversation has accumulated irrelevant context, while supplying the essential facts again.

Your first week should produce a personal cost notebook, not an impressive token total. Record what useful tasks actually cost, adjust your daily ceiling and reserve larger jobs for deliberate spending decisions. Predictability comes from bounded work and visible receipts—not from treating pay-as-you-go access as unlimited.