What one run costs
A typical run of this task sends about 20K tokens and gets back 1.5K, over a single call — 21.5K tokens in total. Those are the numbers below, against the published prices.
| Model | Vendor | Per run | Can do what this needs? | Price evidence |
|---|---|---|---|---|
| GPT-5 Nano | OpenAI | $0.00160 | not stated | Vendor page |
| Gemini 2.5 Flash Lite | $0.00260 | not stated | Vendor page | |
| Claude Haiku 5.5 | Anthropic | $0.00275 | confirmed | Vendor page |
| GPT-6 Luna | OpenAI | $0.00275 | not stated | Vendor page |
| GPT-5.6 Luna | OpenAI | $0.00580 | not stated | Vendor page |
| GPT-5.4 Nano | OpenAI | $0.00588 | not stated | Vendor page |
| Gemini 3.1 Flash Lite | $0.00725 | not stated | Vendor page | |
| GPT-5 Mini | OpenAI | $0.00800 | not stated | Vendor page |
One call, a cent or two, and cheap enough to run on every change rather than the ones you feel nervous about.
Give it the diff and the files the diff touches. A review of a diff alone will flag things that are handled three lines above the window.
It is good at the boring half — unhandled errors, off-by-one, the thing left half-done — and poor at the half that needs to know what the code is for. Treat it as a first pass that frees a human to look at the design.
What we would pick
This section is our judgement, not a figure read off a page. Everything above is arithmetic on published prices; this is an opinion, and it is labelled as one.