Every AI coding tool now has a usage dashboard. Most show the same thing: tokens per user, per model, per day. They tell you how much you spent. They do not tell you what you got for it.
The short answer
Measure AI coding cost per accepted change: attribute every model call, from the platform and from developers’ coding tools, to the story it served, then divide by the changes that were accepted. It shows waste such as abandoned runs and rework, which per-user dashboards hide.
Why per-user spend misleads
The developer with the highest token bill might be your most productive one, or the one stuck in a loop of regenerating the same failing change. Per-user and per-model views cannot tell those apart, because they do not know what the tokens were for.
Attribute every call to the work
The fix is attribution. Every model call is tagged with the story it served: drafting the spec, retrieving context, writing code, running review. That includes calls from developers’ own coding tools, not only from the platform, or half the spend stays invisible.
Tag at the source
Each call carries the story or thread id, the stage and the model used.
Count tokens and price them
Input, output and cached tokens, at the rate of the model and provider used.
Roll up by story
Spec, build, review and rework, so you see where in the lifecycle the cost sits.
Join to outcomes
Mark which stories were accepted, abandoned or reworked.
Cost per accepted change
Divide the total cost by the number of changes someone accepted. Abandoned runs, discarded drafts and rework all count in the numerator, which is the point: they are real spend that produced nothing.
The cheapest token is the one spent on a spec that stopped the wrong build.
| What the breakdown reveals | What you see | What to try |
|---|---|---|
| High rework share | Many tokens on second and third attempts | Tighter acceptance criteria before the build |
| High context share | Large prompts on every call | Better retrieval; fewer, more relevant files |
| Expensive model everywhere | Top-tier model for routine steps | Route simple stages to a smaller model |
| Many abandoned runs | Spend with no accepted outcome | Look at the specs those runs started from |
Watch the trend, not the total
Model prices keep falling, so total spend is a poor signal on its own. Cost per accepted change, tracked month over month, tells you whether the way you work with agents is getting better, independent of what the providers charge.
In CrewPE
We measure proof-of-value engagements against a documented baseline, including cost per accepted outcome. Usage from your AI coding tools, matched to your work, is rolling out.
See how CrewPE worksQuestions teams ask
- How should we measure the cost of AI coding tools?
- Attribute every model call to the piece of work it served, including calls from developers’ coding tools, and divide total cost by the number of changes that were accepted. Per-user or per-model totals alone do not show value.
- What is cost per accepted change?
- The total AI spend for a period, including abandoned runs and rework, divided by the number of changes a person accepted in that period. It rises with waste and falls as specs and retrieval improve.
- How can teams reduce AI coding costs?
- Cut rework with clearer acceptance criteria, send less but more relevant context through better retrieval, and route routine steps to smaller models. Track cost per accepted change to see whether each change helped.
Sneha
AI engineer at EnWithAI, working on CrewPE’s retrieval, grounding and agent evaluation.