Token intelligence
Every AI dollar, and what it actually bought.
AI spend, this month
$48,214
- Claude Code$21,480
- Cursor$12,905
- Copilot$7,640
- Agents (API)$4,312
- Everything else$1,877
Then the part no invoice contains: what each of those dollars cost you to keep.
Your provider can tell you what you spent. Nobody but you can tell you what it was worth, because the return shows up somewhere the invoice never looks: in what the work cost to keep.
Three methods, so every dollar lands somewhere.
Some spend is clean: a provider API gives you the user, the model and the cost. Most is not. Seat licences, shared keys and background agents all resist attribution, and a tool that only reports the easy half leaves the biggest line item unexplained. We say which method produced each figure and how confident it is.
- Verified: native provider API, user and model resolved
- Inferred: commit and session metadata joined to spend windows
- Modelled: time-series allocation where nothing else is available
- Every figure carries its method and a confidence band
- Verified64%
- Inferred27%
- Modelled9%
No bucket called “other”. If we cannot attribute a dollar we say which method failed and why.
Every agent run, and what survived contact with review.
Background agents are the fastest-growing line on the bill and the least legible. A run that opens a pull request nobody merges still costs full price. We record what each run touched, what it cost, and what happened to the output afterwards, which turns agent spend from a mystery into a unit economic.
- Cost, duration and token count per run
- What the run touched, and what it opened
- Merge, revert and rewrite outcomes, followed for ninety days
- Cost per surviving change, by agent and by task type
ROI that survives someone checking it.
The honest version of return is not spend divided by pull requests. It is spend set against what the resulting work was worth, minus what it cost to keep, and the second term is where most of the answer lives. This is the same Slop Index from engineering intelligence, pointed at the bill.
- Spend against surviving work, not merged work
- Net of the rework and review burden the spend created
- By team, initiative, repository and model
- Exportable for finance, with the method stated
- AI spend−$144,642
- Work delivered+$1,930,000
- Cost to keep it−$412,000
- Net$1.37M
A 9.5× return, and a third of the gross eaten by work that should not have merged.