Free tool · runs in your browser
Harness2 cost estimator
Google Research's Harness2 spends money before you see a result: every task schedules executor rollouts and improver sessions that multiply with the budget k. Enter your task count, k and mode and this page shows the exact call counts, then adds your own prices for a floor. The counts come from the repository's documented cost structure and match what the demo logged when I ran it at commit 1cfe254e.
Counts are exact, prices are yours
Nothing here is a price quote. The page never assumes what a model call costs, and nothing you enter leaves your browser.
Checked October 21, 2026
Formulas verified against the demo's logged counts. Re-check the repository before a large run.
I do not fill in prices for you. They depend on the models you point at each role and on how long a rollout runs, so use an average from a small real run.
| Call type | Per task | Whole run |
|---|---|---|
| Executor rollouts4k | 4 | 4 |
| Improver sessions3k | 3 | 3 |
| Selection callsskipped at k of 1 | 0 | 0 |
| Reflection callsone per mode | 1 | 1 |
Executor side
Add a price
4 rollouts
Editor side
Add a price
4 sessions and calls
Cost floor, before judge calls and retries
Enter both prices
Selection and reflection are each priced as one improver session. The repository records them outside its ledger and gives no separate price.
Read this before trusting a total
What the estimate leaves out
Counts are exact for the structure the repository documents. Dollar totals are only as good as the two prices you enter.
Judge calls and retries are not included, so treat the result as a floor and not a quote.
One rollout can run for a long time. A price per rollout is an average you have to choose, and a single expensive rollout can move the total.
The formulas were checked against the demo at the commit above. A later release may change the structure, so re-check the repository before a large run.
Questions
About the estimator
k is the budget for each task. In one mode a task schedules k base executions, k invariant candidate runs, k domain candidate runs and k final runs, which is 4k executor rollouts. The improver side is k invariant proposals, k domain proposals and k compositions, which is 3k sessions. Doubling k doubles both numbers.
Related
Pointing an agent at your own tasks?
I build and tune agents that have to hold up on real work, with the spend and the failure modes worked out before launch. If you want a second pair of eyes on a setup like this, book a call.
Book Meeting