Question 1

How much does AI code review cost per pull request?

Accepted Answer

For a typical pull request on a mid-tier model the token cost is somewhere between one and five cents, and the number is driven far more by how you assemble the prompt than by the size of the diff. A few hundred changed lines is only a couple of thousand tokens; the system prompt, coding guidelines and surrounding repository context that get attached to every review are usually larger than the diff itself, and they are sent again on every follow-up pass. Output is the other half of the bill, because review comments are long-form text priced at four or five times the input rate, so a bot that writes an essay on every file costs multiples of one that writes three targeted comments. Multiply by the number of passes — most pull requests get re-reviewed at least once after changes — and a team pushing several hundred pull requests a month typically lands in the tens of dollars, not the hundreds. That is small enough that the interesting question is not the token bill but whether the reviews are worth reading.

Question 2

Does prompt caching actually help for code review?

Accepted Answer

More than for almost any other workload, because code review is the rare case where a large block of the prompt is genuinely identical every time. Your review instructions, style guide, architecture notes and any pinned files do not change between pull requests, and cached reads are typically billed at around a tenth of the normal input rate. If that fixed block is the majority of your input tokens — which it usually is once you attach real repository context — caching removes most of the input side of the bill. The catch is ordering and lifetime: the cached block has to come first in the prompt and has to be reused before the cache entry expires, so a repository with a trickle of pull requests spread across the day may miss the window and pay full price anyway. The calculator lets you set the cached share directly so you can see the difference between a busy monorepo that hits cache constantly and a quiet one that rarely does.

Question 3

What actually makes AI code review expensive?

Accepted Answer

Not the tokens — the false positives. A bot that flags one or two things per pull request that turn out to be non-issues costs a couple of minutes of a reviewer's attention each time, and a reviewer's minute is worth vastly more than a thousand tokens. At any realistic salary, two minutes of dismissing noise costs more than a hundred reviews' worth of API calls. That is why this calculator prices the triage time alongside the token bill and subtracts both from the reviewer time saved: it is the only way to see that tuning the bot to be quieter is worth far more than switching to a cheaper model. The same logic explains why teams abandon these tools despite a trivial invoice — the invoice was never the problem. If the calculator shows a negative net, the fix is fewer, better comments, not a smaller model.

Question 4

Is a per-seat code review tool cheaper than calling the API yourself?

Accepted Answer

It depends almost entirely on pull requests per developer rather than on team size. Hosted review tools charge per seat per month, so their cost scales with headcount and is flat in usage; calling the API yourself scales with pull requests and is flat in headcount. A team where each developer opens a handful of pull requests a month will find raw API calls dramatically cheaper, because the token cost per pull request is cents while a seat is tens of dollars. A team shipping many small pull requests per developer per day narrows the gap, and at that point the hosted tool's integration, inline comments, and lack of maintenance burden are usually worth the difference. Use this calculator to get your true per-pull-request API cost, multiply by your volume, and compare it against the seat quote — and remember to add the engineering time to build and maintain the integration, which is the cost most build-versus-buy comparisons quietly omit.

AI Code Review Cost Calculator

Where the tokens go

The same reviews on different model tiers

Tune the noise before you tune the model

この計算機の仕組み