Measure what the provider charges.
Input, output, reasoning, cache reads, cache writes, batch pricing and retries affect the bill differently. Capyba keeps provider accounting categories separate and records gaps when usage data is incomplete.
Find changes worth testing.
Capyba ranks recoverable waste using the available usage records, confidence and overlap between opportunities. Repeated context, duplicate instructions and stale knowledge are possible contributors. Focus on changes whose likely value justifies their cost, rather than the largest expense alone.
Protect the rules the team relies on.
Review instruction and memory changes as a diff. Preserve protected rules, require the right approval and retain a rollback path.
Judge quality for the task.
An accepted task meets explicit criteria for the work. Code needs relevant tests and review. Text needs supported claims, useful output and acceptance by the person doing the work. Fewer tokens are not a saving if the task needs more correction.
Track the evidence, from opportunity to outcome.
| Evidence state | What it means |
|---|---|
| Identified | A possible opportunity has been found. |
| Simulated | Its effect has been estimated using stated assumptions. |
| Validated | A comparable experiment passed the acceptance checks. |
| Realised | Observed eligible usage or billing supports the saving. |
| Unknown | The available evidence cannot support a result. |
| Rejected | The change failed the required checks. |
Measure eligible spend.
A tested change may reduce the bill. Include model charges, cache, retries, tools, connectors, indexing and the allocated cost of Capyba.
Measure usable capacity.
A change may yield more capacity or fewer interruptions without lowering your subscription fee. Report the quantity you can actually observe.