Code review for $2: a beginner's token receipt

01 / Set a 2 dollar ceiling before the first review request

Most beginners lose money on AI code review because they start with the repository instead of the arithmetic. The fix is boring and effective: pick your ceiling first. Decide that the entire first week costs no more than 2 dollars, and treat every request as one line on a receipt. When the receipt reaches the ceiling, you stop. That single habit teaches you more about prompt design than any free trial does, because free trials hide the cost of sloppy prompts.

CheapRelay keeps that arithmetic honest. There is no monthly fee and no subscription, and billing is per real token, including cache reads and cache writes. Nothing sits in a plan bucket waiting to expire. One key reaches Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more, and the same key works in Claude Code, Codex, Gemini CLI or any OpenAI SDK. The API base is https://api.cheapapirelay.com, with OpenAI-compatible routes under /v1 and Gemini native routes under /v1beta, so the setup you learn once transfers to other tools.

Two habits come before any code. First, write down your ceiling in your notes so it is visible while you work. Second, decide what a review must produce: a short list of concrete defects with file and line references. A review that returns prose is a review you paid for twice, once for the tokens and once for the rereading.

02 / Start with one low-cost model and one slice of the diff

Do not open with the largest model you can reach. For a first pass, Claude Haiku 4.5 is a sensible choice: it is on sale, it is fast enough for a loop of small reviews, and a cheap first pass usually catches the boring defects before a more expensive model sees them. If you want a second opinion on a risky change, DeepSeek V4 Flash is another on-sale option to compare against. Two models on one change is a reasonable ceiling for a beginner; three is where the receipt starts to wander.

The slice matters as much as the model. A 400-line diff is a review unit. A 6,000-line branch is not a review unit, it is a research project, and its token count behaves like one. Split work by feature or by commit, and review each slice on its own. If a slice cannot be described in one sentence, it is too big to review well and too big to review cheaply.

Keep the scope narrow in the prompt as well. Ask for defects, not opinions on architecture. Ask for file and line references, not a rewritten file. Ask for a severity label of blocking, should-fix, or nitpick, so you can triage without rereading the answer.

03 / Shape the prompt so the diff goes in and findings come out

A budget review prompt has three parts and no small talk. The first part states the rules: report defects only, cite file and line, cap the output at the number of findings you actually need. The second part is the diff itself, pasted as plain text with enough surrounding context to resolve symbols. The third part is the format you want back.

• ■ State the change intent in one sentence so the model does not guess and then argue with its own guess.
• ■ Include only the files touched by the change, plus the signatures of anything they call.
• ■ Ask for at most 10 findings. A cap forces ranking, and ranking is what you are paying for.
• ■ Ask for a one-line fix suggestion per finding instead of full code, which cuts output tokens sharply.
• ■ Reject pasted stack traces, lockfiles, and generated code. They are noise priced at the same rate as signal.
• ■ Use the same instructions every time so cache reads can do the work on repeat runs.

The last point is the quiet saver. The instructions rarely change between reviews, so if your tooling sends them as a stable prefix, the second and later runs can read them from cache instead of reprocessing them. CheapRelay bills cache reads and cache writes as real tokens on the same ledger, so you see exactly what the reuse costs and what it saves. You are not gambling on an opaque plan; you are watching two numbers instead of one.

04 / A sample evening receipt: just under 50,000 tokens

Here is a realistic first session for a beginner, sized in tokens rather than in feelings. Three reviews, each on a diff of roughly 400 lines.

• ■ Review 1: about 8,500 tokens of diff, 700 tokens of standing instructions, 900 tokens of findings output.
• ■ Review 2: about 8,000 tokens of diff, 700 tokens of instructions, 850 tokens of output.
• ■ Review 3: about 7,600 tokens of diff, 700 tokens of instructions, 950 tokens of output.
• ■ Cache activity across the three runs: roughly 3,000 tokens written and 12,000 tokens read.
• ■ Rough total: about 48,000 to 50,000 tokens for the evening.

That total is the honest unit of cost here. Multiply it by the per-token rate shown for the model you chose, add nothing else, and you have the receipt. Compare it with the same evening done badly: the whole repository in the prompt, no output cap, three models asked the same question. In my experience that version runs into the hundreds of thousands of tokens for roughly the same three findings, because most of the spend went to text the model did not need to see.

Keep a running log with four columns: date, request size, output size, and what you found. After a week you will know your personal cost per real defect, and that number is the only benchmark that matters. When the log shows that small diffs catch most issues, you will stop sending large ones on your own.

05 / Cache reads, retries, and four habits that cut the bill

Cost control is mostly retry control. Each extra retry is a full request priced at the same rate as the first, and retries pile up when the prompt is vague. Tighten the prompt instead of resending it.

• ■ Batch small questions. One request with three questions beats three requests with one each.
• ■ Never resubmit an unchanged prompt to force a different answer. Edit the prompt to say what was missing.
• ■ Trim the diff before you send it. Removing one generated file can save more tokens than an hour of clever phrasing.
• ■ Cap output length. Ask for 10 findings and 3 sentences each, and stop the model from narrating its process.

One more thing worth knowing before you commit to a tool: CheapRelay pays a referral reward of 20 percent of what invited users pay, which is a small offset if your team joins you later. It does not change your per-token rate, so it should never be the reason you pick a workflow. Sign up on CheapRelay, load the 2 dollars you set as your ceiling, point Claude Code or Codex at the API base, and run one 400-line review. If the receipt looks like the sample above, keep the habit. If it does not, shrink the diff before you change models.

Budget code review is not about finding the cheapest model and hoping. It is about sending less text, capping the output, reusing the stable prefix through cache reads, and reading the receipt after every session. Do that for two weeks and you will spend a few dollars to buy the skill that expensive teams pay salaries for: knowing which change deserves a second pair of eyes and which one does not.