
Cost per task: the real unit of AI spend
In short: Cost per task measures the full cost of finishing one unit of work with AI, including retries, tool calls and human review, not just token price.
Cost per task is the total cost of completing one unit of useful work with an AI system, such as one resolved ticket, one reviewed contract or one merged pull request, including every model call, retry, tool call, retrieval step, infrastructure share and human review needed to get there. It is a better decision metric than price per token, because cheap tokens used badly can make an expensive task.
Cost per task in plain words
The price of petrol per litre tells you little about what it costs to get from home to the office, detours included.
How it works
- Define the task and what "done" means.
- Trace a sample of real tasks end to end: tokens in and out per call, number of calls, tools used, failures and retries.
- Add infrastructure and the cost of human time for review or fixes.
- Divide by successfully completed tasks, not attempts.
Example
As an illustration: model A costs a third as much per token as model B but needs three agent loops and fails a quarter of the time, triggering human rework. Model B finishes in one pass with fewer failures. Per completed task, B can come out cheaper.
Where you'll meet it
- Agent platform dashboards and AI gateway reports
- Business cases and ROI models
- Model routing decisions
- Vendor comparisons, where per-token rates alone mislead
What it changes for you
For finance and product leaders, it ties AI spend to business outcomes and makes pricing your own AI features possible. For engineers, it rewards fewer loops, shorter prompts, caching and good routing. For buyers, it is the only fair way to compare vendors with different pricing models.
Cost per task vs token pricing
Token pricing is the provider's rate per million tokens, as listed on OpenAI's and Anthropic's pricing pages. Cost per task is what you actually pay to get one job done, including everything around the tokens.
Common mistake
My view: dividing by attempts instead of successes. A cheap attempt that fails half the time doubles the real cost. Count only work that met the bar.
Related terms
Token pricing · Model routing · AI ROI · Prompt caching · Agent loop
From the archive: Price AI work by accepted outcome, not by token
Divide spend by tasks that met the bar, never by attempts.


