The economics of AI code review: what a review actually costs

AI code review has an unusual cost structure for SaaS: the marginal cost of serving a customer is real, variable, and occasionally large. Most software costs approximately nothing per additional use. This does not.

Two real reviews

A 17-file pull request in Standard mode: one chunk, about 50,000 input tokens, 4,300 output tokens, roughly a minute, a few cents of AI spend.

An 80-file pull request in Deep mode: four chunks, each reviewed twice, 894,000 input tokens, 113,000 output tokens, around thirty minutes, and several dollars of AI spend.

That is a difference of two orders of magnitude between two reviews on the same plan.

Where the tokens go

Not mostly on the diff. A useful review needs the surrounding code the change touches, and that context is re-sent with every chunk. Deep mode adds a second pass over each chunk, which roughly doubles it again.

Prompt caching helps considerably within a single review, because the shared context is identical across chunks. It does nothing across reviews — the cache expires in minutes, so re-reviewing the same pull request an hour later costs full price.

Why this makes flat pricing awkward

If you charge one price per developer per month, you are averaging across customers whose usage differs by 100x. That works while your customer base is homogeneous, and stops working the moment one customer ships a monorepo refactor every week.

The industry answer so far has been to charge per seat and quietly rate-limit the heavy users. We would rather meter honestly and publish the ceilings: 10 units for Quick, 20 for Standard, 40 for Deep, whatever the size of the diff.

What we do about the cost

Quick mode uses a smaller model. Incremental re-reviews only look at commits since the last reviewed SHA rather than re-reading the whole pull request. Findings below the confidence threshold are discarded rather than published, which does not save tokens but does save your attention — which is the scarcer resource.