On this page
The short answer
Attribute AI cost to the complete user task, linking every model attempt and supporting operation to its intended outcome. Separate provider-reported usage, local estimates, and unknown amounts. Cost per response can hide retries, failed work, retrieval, and evaluation that determine whether a cheaper model actually reduces the cost of useful completion.
What to take away
- Keep task and attempt identifiers distinct.
- Retain failed and canceled work where it consumed resources.
- Report unknown cost and zero-success cases explicitly.
Sources checked . Research synthesis and Rubrex recommendations. Examples and local experiments are labeled; they are not client results.
Choose the decision the cost report supports
Langfuse documents model usage and cost views. Rubrex recommends defining an application cost ledger that can answer a concrete decision, such as whether a route reduces cost per accepted support task. The ledger should not silently equate displayed model tokens with the entire operating cost.
Use a synthetic support task with one retrieval, two model attempts, a judge call, and a final outcome label. Give each operation a stable task reference and usage provenance. Keep price inputs dated and external to the illustrative calculation; this article supplies no current vendor rate.
Evidence: Langfuse: Observability and Application Tracing [1]
Link attempts and supporting work
Record requested model, actual destination, attempt reason, operation timing, usage source, and final task status. Include a failed attempt that consumed resources and a cache hit whose cost behavior follows the chosen accounting contract. Avoid charging a retry as an independent successful user task merely because it created a new trace.
Add retrieval and evaluation as separate ledger categories when they are in scope. State whether infrastructure and human review are excluded or estimated rather than presenting a partial model bill as total cost. The relevant boundary depends on the decision, but it must remain consistent between compared routes.
Reconcile delayed, missing, and duplicated usage
Deliver one usage event late and another twice. The ledger should update the intended operation once without losing its relationship to the original task. A missing usage response remains unknown or estimated under the stated policy; converting it to zero can make the least observable route appear cheapest.
Compare the task ledger with the authorized provider or gateway accounting export where available. Keep differences visible and investigate rounding, timing, and scope before declaring an error. Include canceled work and partial tasks so the report reflects resources consumed even when the user received no accepted result.
| Case | Evidence to retain | Acceptance condition |
|---|---|---|
| Failed paid attempt | Task-linked usage | Included once in scope |
| Duplicate usage event | Operation identifier | No double charge |
| No accepted tasks | Outcome denominator | Accepted-task cost undefined |
Compare accepted-task cost with its limits attached
For the synthetic comparison, divide total in-scope measured and estimated cost by tasks meeting the declared acceptance criteria, reporting unknown amounts separately. If no task passes, cost per accepted task is undefined rather than zero. Retain quality, latency, and critical failures beside the cost figure.
Accept the attribution pipeline when every fixture operation is linked once, late usage reconciles, and exclusions remain explicit. Re-run after changing routing, caching, or evaluation. The resulting ledger supports a bounded trade-off decision; it is not a guaranteed invoice forecast or proof that lower token prices produce better business economics.
Limits of the evidence
The task and ledger events are synthetic; no billing audit or price comparison was performed. Actual rates, discounts, rounding, and charge timing depend on provider terms and deployment. Cost attribution does not establish task quality or financial return by itself.
Common questions
Should failed requests be free in the report?
Include resources actually consumed under the chosen accounting scope, even when the task failed.
Can model-token cost represent total cost?
Only label it as that component; retrieval, evaluation, infrastructure, and review may add other costs.
Sources & further reading
Primary sources behind this briefing. A source’s findings apply to its own study conditions; publication on arXiv does not establish peer review.
- Langfuse: Observability and Application Tracing Langfuse · Living reference · Maintained documentationReviewed: documented tracing, scores, usage views, and experiment workflows; no hands-on comparison. Accessed October 16, 2026.
Questions or corrections? Write to Rubrex. Read our editorial approach.