One $5 Receipt Covers Claude, Codex and Gemini CLI

01 / The desktop budget is $0 and the first API cap is $5

Most AI desktop write-ups open with a shopping list. This one opens with a receipt you will not have: no new computer, no annual plan, no bundled seat. A command-line client is a thin piece of software. The model runs on a remote endpoint, so the machine in front of you only has to render text, read files and write edits back to disk.

• ■ A laptop you already own, even a five-year-old one.
• ■ 8 GB of RAM is comfortable, 4 GB works if you close the browser tabs.
• ■ Node.js version 20 or newer, which all three CLIs expect.
• ■ About 2 GB of free disk for dependencies, transcripts and session logs.
• ■ Any operating system: Linux, macOS, or Windows with WSL for the smoothest path.

That leaves exactly one budget line to worry about: API spend. Cap it at $5 for the first month. Five dollars is a deliberately small experiment rather than a price quote. The point is that every wrong turn, every abandoned session and every silly question costs pennies while you work out which tool you actually reach for. If you already have an old laptop and a working internet connection, your cash outlay before the first token is zero.

02 / One key wires Claude Code, Codex and Gemini CLI together

The three clients have separate installers and separate config files, but they can all speak to the same relay. CheapRelay is a pay-as-you-go AI API relay where one key reaches Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more. That matters on a small budget: instead of juggling three provider accounts and three invoices while you compare them, you keep one key and one usage page.

Install order hardly matters. Get the clients working first, then hand each of them the same two facts: a base URL and the key. CheapRelay exposes OpenAI-compatible routes under /v1 and Gemini native routes under /v1beta, both rooted at https://api.cheapapirelay.com.

• ■ Claude Code: install the CLI, point its base URL at the relay, paste the key into its auth setting.
• ■ Codex: install the CLI, set the OpenAI base URL to https://api.cheapapirelay.com/v1, use the same key.
• ■ Gemini CLI: set the base URL to the relay's /v1beta path for Gemini, use the same key again.
• ■ Export the key once as an environment variable so it never gets pasted into a chat or committed to a repo.
• ■ Send one trivial prompt through each client before you touch a real project, and confirm a reply arrives in all three.

Sign up on CheapRelay, generate one key, and the billing setup is done. Billing is per real token with no monthly fee and no subscription, so a month where you barely touch the keyboard costs almost nothing. There is nothing to cancel when life gets busy.

03 / A worked receipt line: 214,000 tokens in one evening

Abstract advice about token prices is useless to a beginner. Here is a concrete receipt line from an evening of refactoring a small Node project, rounded so the arithmetic is easy to follow. Picture forty turns of conversation, where every turn resends the file you are editing plus the recent history.

• ■ Fresh prompt tokens: about 35,000.
• ■ Cache reads: about 140,000, the same context resent on later turns.
• ■ Cache writes: about 25,000, the first time that context was stored.
• ■ Output tokens: about 9,000, the assistant's answers and file edits.
• ■ Counted total: 214,000 tokens, because all four of those lines are billed.

That is the honest shape of an AI bill, and the surprise is line two. Roughly two thirds of the counted tokens in that session were cache reads, not fresh thinking. CheapRelay bills per real token including cache reads and writes, so those repeats show up on the receipt instead of being quietly discounted away. Knowing that early is what keeps a $5 cap intact: the fastest way to burn money is not asking hard questions, it is sending the same enormous context four hundred times.

04 / Cache reads and writes are the line item beginners miss

Once you accept that context is the product you are paying for, the habits write themselves. You are not trying to be clever with prompts; you are trying to keep the payload small and stable so cache reads stay cheap in volume and cache writes happen once instead of twenty times. This is the difference between a $5 month and a $5 weekend.

• ■ Restart the session instead of letting one conversation run all night. A fresh session with a tight prompt beats a 200-turn thread carrying stale noise.
• ■ Use the client's compact or summarise command when the transcript grows past what you actually need.
• ■ Never paste a whole repository. Paste the file being changed and two or three relevant neighbours.
• ■ Ask for diffs and targeted edits rather than full rewritten files, since output tokens are the priciest line in the receipt.
• ■ Route grunt work to the smaller models on sale, such as Claude Haiku 4.5 for quick edits and Gemini 2.5 Flash for summaries, and save the heavyweight model for the one hard design decision of the day.
• ■ Check the usage page after the first real session. Ten minutes of reading your own receipt teaches more than any pricing blog post.

A small table beside the keyboard helps too. Keep a running note of how many tokens each session consumed and what it was for. After a week you will know which of the three CLIs earns its place on your machine, and that is worth more than any feature comparison.

05 / Stretching five dollars across four weeks of tinkering

Here is a simple way to pace the month so you end it with money left and something built. Treat the $5 as four buckets, not one pile, and stop when a bucket empties rather than borrowing from the next.

• ■ Week 1: $1 of setup and probing. Three installs, a handful of throwaway prompts, and a first honest receipt reading.
• ■ Week 2: $1.50 on a small real task, like adding tests to an existing file, using one CLI only.
• ■ Week 3: $1.50 repeating the same task in a second CLI so you compare like with like.
• ■ Week 4: $1 of cleanup and documentation, which is where cheap models shine.

Two frugal closing moves. First, pay for what you use and nothing else: since there is no subscription, an unused week is genuinely free, so let quiet weeks stay quiet. Second, the referral reward pays 20 percent of what invited users spend, so if a friend starts with their own $5 experiment, a slice of their usage comes back to you and stretches your own cap without any extra work.

The whole setup comes down to one laptop, one key, three thin clients and a hard spending ceiling. Skip the hardware, keep the context small, read your receipt after every session, and the first month costs less than a single takeaway coffee. After that you will know whether this workflow deserves a bigger number or exactly the same one.