# What is Cost per action?

Also called unit cost, cost per task, cost per run.

Cost per action is the total resource cost attributed to one completed unit of work, such as one answered ticket, one drafted document, or one enriched record. It aggregates model tokens, tool and API calls, storage, retries, and failed attempts. Because a single action often involves many model calls, cost per action is usually far higher than the cost of one call.

The calculation starts from a stable identifier for one unit of work, then sums every metered event carrying that identifier: model tokens across all steps, third-party API calls, storage and retrieval, and any compute time. Retries and abandoned attempts are attributed to the same unit, because they were spent pursuing that outcome. Shared costs such as infrastructure are amortized rather than ignored.

The number matters because it decides whether an automation is viable at all, and because agent workloads have unusually wide variance. It informs pricing, capacity limits, and where engineering effort pays back. It also makes regressions visible: a prompt change that adds one extra step per run may be invisible in quality metrics and obvious in cost per action within a day.

Two errors are common. The first is measuring only successful runs, which hides the cost of everything that failed, was retried, or was abandoned by the user. The second is reporting a single average across task types that differ by an order of magnitude. Because the distribution has a long tail, a median plus a high percentile describes the workload far better than a mean.

Cost per action sits downstream of token usage and retry behavior and upstream of pricing and capacity decisions. It also connects to rate limits and concurrency, since the practical way to bound spend from a runaway loop is to cap the number of steps and the amount of parallel work a single unit is allowed to consume.

## Key points

- Aggregates every call, retry, and failure behind one unit of work.
- Failed and abandoned attempts belong in the numerator.
- Long-tailed distributions mean medians and averages both mislead alone.
- Requires a run identifier that ties spend to an outcome.

## In practice

A team measures the cost of resolving one inbound support email. The median run uses three model calls and one knowledge search. But one run in twenty escalates, loops through eight steps, retries a failing tool twice, and then gets discarded. Counting only median successes understates true cost by roughly a third, which changes whether the automation is worth keeping.

## Related terms

- [Token usage](/en/glossary/token-usage)
- [Retry policy](/en/glossary/retry-policy)
- [Rate limit](/en/glossary/rate-limit)
- [Observability](/en/glossary/observability)
- [LLMOps](/en/glossary/llmops)

[Back to the AI Glossary](/en/glossary)