How the AI work cost index is measured

What is counted

One workspace: the one where we build Hourtick. In it, AI agents are members. They are handed tasks, work on them through Hourtick's MCP server and report back. Each piece of work is a session with a start, an end and a result.

The index counts sessions that finished as done in the last 30 days, where the agent reported both working time and cost. Sessions that failed or were cancelled are not in the cost numbers; their share is shown separately.

Where the numbers come from

• Time is the seconds the agent spent working, summed over its working periods. Time spent waiting for a person's answer is not counted.
• Cost is what the agent reported for the session in US dollars, from the tokens it used and its provider's price. Hourtick does not run the models, so this is the agent's own number. Where a provider's bill is connected, the workspace's totals are reconciled against it.
• Tokens are input plus output tokens as reported.
• The model is the name the agent reported. One agent can use several models in a session; the session is listed under the one it reported.

Why medians

Agent work has a long tail: most tasks are cheap and a few are very expensive. An average is pulled up by those few, so the index shows the median (half of the sessions cost less) and the 90th percentile (nine in ten cost less).

A group is shown only when at least 5 sessions stand behind it. Cost per hour is total cost divided by total working time for the group.

What it does not include

• The person's time: writing the task, answering questions, reviewing and fixing. For most tasks that is the larger cost. The cases describe it; the index does not price it.
• Subscriptions. Where an agent runs on a flat-rate plan, its reported cost is what the same tokens would have cost at list price, when the agent reports that at all.
• Quality. A cheap session that needed a second attempt is counted as two sessions, not as one expensive one.

What it can and cannot tell you

It is one team's software work: features, fixes, reviews, research and writing on a TypeScript codebase. It is useful as a measured reference point and for seeing how cost moves when models and prices change. It is not a market average, and your tasks will differ.

No task titles, people or client names are published in the index. The cases are about our own product work and are written up separately, each with its own numbers.

Using the numbers

You may quote the index with a link to this page and the date you read it. If you find something that looks wrong, tell us through the contact page; corrections are made in the open.