What an AI Code Review Actually Costs Now

The pricing

GPT-5.6 Luna now costs US$0.20 per million input tokens and US$1.20 per million output tokens. A code review sends the pull request diff plus a prompt as input and gets the review text back as output. That is the entire marginal cost: Cody, our review tool, is a GitHub Action that runs on your own API key, so what OpenAI charges is what a review costs.

Three real pull requests

These are public pull requests on the Cody repository itself, reviewed by Cody as part of our own workflow. You can read every review comment. Token counts are estimated from the actual diff and review sizes at roughly four characters per token, plus prompt overhead, priced at Luna's current rates.

1. PR #20, a small change (+132/-3): about 3,000 tokens in, 500 out. Cost: about $0.001.
2. PR #12, a typical feature (+179/-25) that produced a detailed 10,000-character review: about 4,700 tokens in, 2,500 out. Cost: about $0.004.
3. PR #14, a large change (+1,840/-203): about 37,000 tokens in, 750 out. Cost: about $0.008.

Estimates, not invoices: exact token counts vary with diff content and the model's response length. The order of magnitude is the point.

What that adds up to

A dollar covers roughly 200 to 800 reviews depending on pull request size. A team merging 100 pull requests a month spends about 50 cents on review inference. Our own monitoring runs a real review every day and the OpenAI usage dashboard rounds the month to $0.00.

The practical consequence: there is no longer a budget conversation to have about AI review on every pull request. The cost question moved from "per seat per month" to "less than the coffee round, annually".

Using GitLab or Azure DevOps? Cody Pro runs the same reviews in GitLab CI and Azure DevOps pipelines.