Skip to content

Measure what a run costs, and publish the numbers #5

Description

@RobYed

docs/costs.md describes what drives the cost of a run and says honestly that the numbers are not measured. That is the right thing to write and the wrong thing to leave standing: for a team deciding whether to adopt this, the difference between "interesting" and "too expensive" is a number.

What to build

The data already exists per run: the transcript prints tokens and, when the endpoint publishes prices, the cost. It has never been collected.

  • Aggregate a few dozen real runs per agent: input tokens, output tokens, wall time, and the size of the diff they were given.
  • Publish medians and a rough upper bound in docs/costs.md, with the model and the provider named, and dated. A number without those three is worse than none.
  • Separately for the acceptance test, where the number of turns dominates and the variance is much larger.

Perhaps also

A cost ceiling. Nothing in the package stops a run from being expensive; the timeouts are the only backstop, and they measure time rather than money.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions