JMLR Software
I ship real software with AI agents. Here is what it costs.
Ten years as a software engineer, now running client work and my own products through an agent pipeline. I write down what actually broke and what it actually cost, with the receipts.
Send it to meWhat is your coding agent actually costing you?
Most estimates are guesses. These numbers come from two real runs I measured on my own client repositories, so you can see the sample size and decide whether to trust it.
One run is one task handed to the agent, start to finish.
A turn is one model call. A small change is about 30.
$2.00 in / $10 out per million tokens, read 2026-06-24.
Per month
$24
- Per run
- $0.54
- Per turn
- $0.02
- Per year
- $282
What that number is made of
- Cached input$0.0045
- Fresh input$0.0050
- Output$0.0060
Cost tracks turn count, not repository size. A whole client site is under two percent of the context window, so making the repo smaller saves almost nothing. Thirty-five turns for a two-line change is the number worth attacking.
The runs this is calibrated against
- $0.55 over 35 turns
- $0.42 over 32 turns
Two real runs, both Sonnet 5 with prompt caching on by default. That is the entire sample. Every other cost calculator is extrapolating from nothing and does not say so.
Assumptions, stated so you can disagree with them
- About 25,000 input tokens and 600 output tokens per turn.
- Ninety percent of input served from cache, which is the default in the GitHub Action.
- List prices for the Anthropic first-party API. Your platform may differ.
An estimate from a two-run sample, not a bill. Check your own usage console before making a decision on it.
20 ways an agent pipeline actually breaks
Every entry is a real failure with a date: what you see, what caused it, what to do now, and what to change so it does not happen again. Nothing theoretical, and nothing you can get by asking a model. Yours for an email address.
No spam, and one click unsubscribes. I do not sell the list.