What costs money
Four actions spend model budget. This page says which, roughly how much, how to set a limit, and what happens when you reach it.
Rehearsal uses language models to build a world, to play the simulated customer, and to run an agent that you describe with instructions. Those model calls cost money. Four actions make them. Everything else is free.
The four actions
| Action | Typical cost | Typical time | Where |
|---|---|---|---|
| Build a world | $3 to $8 | 1 to 3 hours | Build world, rehearsal build, build_world |
| Run an evaluation | about $0.02 for each episode | 2 to 10 minutes | Start evaluation, rehearsal eval, run_evaluation |
| Improve an agent | under $0.10 | about 1 minute | rehearsal improve, improve_agent |
| Re-check a world | about $0.03 for each job | 5 to 20 minutes | rehearsal worlds validate, recheck_world |
These are the product's own figures at version 0.1.2. The measurements behind them, recorded in the product's code, are $1.70 to $8.84 for a build and $0.016 to $0.018 for an episode. Your application can cost more or less: a larger application takes longer to build.
Reading results, listing worlds, creating agent profiles, comparing runs and exporting data are free.
How many episodes is a run
episodes = jobs in the chosen splits x agents x repeatsA world with 6 dev jobs, 2 agents and 2 repeats needs 24 episodes, which is about $0.50. When you choose no split
and no repeats, a run uses every approved job twice.
Who pays the model provider
- On the hosted service, the model calls run on Rehearsal's model account. They count against your workspace's monthly allowance, which is part of your plan. You do not need a model provider key.
- On a server you run, the calls use the provider keys that the operator configured, and the provider bills the
operator.
rehearsal priceson the server shows the price of each configured model. - With a bring-your-own agent, your agent's own model calls are yours, billed by your provider. The world and the simulated customer still count against the allowance.
Set a limit
You have three limits, from the smallest to the largest.
A budget for one job
Give a build or a run a budget. The job stops itself when it has spent that much.
rehearsal build <app_id> --budget 6
rehearsal eval <wv_id> --profile <profile_id> --split dev --budget 1On a plan with an allowance, a job that you start without a budget gets a default cap, so that one job cannot hold the whole month's allowance: $20 for a build, $3 for a re-check, $2 for an improvement, and $0.06 for each episode of a run. On the Free plan a build stops at $5. Plans and limits has the current numbers.
The monthly allowance
Each plan has a model allowance for the calendar month. It is a safety net: work that would pass it does not start. The allowance covers all four actions.
To see what is left:
Open Billing, or API keys: both show the spend of the month against the allowance.
The plan's counts
A plan also limits how many applications you keep, how many builds you start, and how many episodes you run in a month. See Plans and limits.
On a server you run, there are no plans unless the operator turns them on. The operator can set one spend limit
for the whole server with REHEARSAL_MONTHLY_ALLOWANCE_USD.
At a limit
| What happens | What you see | What to do |
|---|---|---|
| A job reaches its own budget | The job stops. A build fails with "budget" in its error | Start again with a higher budget, if the result is worth it |
| The allowance is used up | Status 402: "This workspace has used its ... model allowance for this month" | Wait for the 1st of the month, or stop a job that is still running. On a paid plan, add a top-up |
| The plan's episodes are used up | Status 402: "This run needs ... episodes and the workspace has ... left this month" | Run fewer: one split, fewer jobs, or one repeat |
| Every sandbox is in use | Status 429 | Wait for a run to finish, or stop one |
Nothing is started and nothing is charged when a request is refused.
What you do not pay for twice
- A failed build can be restarted. A restart continues from the last saved phase. Finished phases are not repeated, and the build is not counted again.
- A build that cannot go on stops early. A phase that meets the same problem several times stops, so a failed build costs less than a full one.
- Episodes that never started are not counted. If you stop a run, or it fails, the episodes it had not started go back to your count.
- An episode that the environment broke is not scored. It does not count against the agent.
Before an assistant spends
An AI assistant that uses Rehearsal must tell you what it will start, how many builds or episodes, and the rough cost, and then wait for a clear yes. The skill that ships with the plugin gives it this rule. The plan's limits are a hard stop whatever the assistant decides.
Was this page helpful?