Plans and limits
What each hosted plan allows, exactly as the server enforces it, and what happens at a limit.
The hosted service puts every workspace on one plan. The numbers on this page come from the same table the server checks before it starts work, so they cannot differ from what is enforced.
Self-hosted servers
A server you run yourself has no plans unless you turn them on with REHEARSAL_PLANS=on. Without plans, the
only spend limit is the server-wide REHEARSAL_MONTHLY_ALLOWANCE_USD, which is off by default.
What each plan allows
| Free | Pro | Team | Enterprise | |
|---|---|---|---|---|
| Applications kept | 1 | 3 | 20 | 100 |
| World builds | 1, once | 2 a month | 8 a month | 40 a month |
| Episodes a month | 40 | 750 | 3,000 | 20,000 |
| Sandboxes at once | 1 | 2 | 4 | 8 |
| Builds in progress at once | 1 | 1 | 2 | 4 |
| Model allowance a month | $8 | $60 | $250 | $1,500 |
| Spend cap of one build | $5 | your budget, or the default | your budget, or the default | your budget, or the default |
| When evaluations start | in quiet hours | at once | at once | at once |
| Top-ups | no | yes | yes | yes |
A month is a calendar month in UTC. Counts start again on the 1st.
What the words mean
| Limit | What counts |
|---|---|
| Applications kept | Applications in the workspace now. Remove one to add another. |
| World builds | Each build you start. A failed build can be restarted from its last saved step and is not counted again. The Free plan's 1 build is for the life of the workspace, not for a month. |
| Episodes | One attempt by one agent at one job. A run needs jobs x agents x repeats episodes, and they count from the moment the run is queued. Episodes that a stopped or failed run never started are not counted. |
| Sandboxes at once | How many parts of your evaluations run at the same time. More sandboxes finish a run sooner. |
| Model allowance | A safety net on model spend for the month. It covers builds, evaluations, re-checks and improvements. Work that would pass it does not start. |
Quiet hours
On the Free plan, an evaluation starts only in the quiet window. The window is 18:00 to 02:00 UTC on the hosted
service's default setting. A run that you start outside the window is accepted and waits: the answer includes
starts_at, the time it will start. Paid plans start at once.
Top-ups
Pro, Team, Enterprise workspaces can add to the current month on the console's Billing page.
| Top-up | Adds | Adds to the model allowance |
|---|---|---|
| 1,000 episodes | 1,000 episodes for this month | $20 |
| 1 world build | 1 world build for this month | $10 |
Default spend caps
A job that you start without a budget gets a cap, so that one job cannot hold the whole month's allowance
while it runs. A job stops itself when it reaches its cap. Give your own cap with --budget or budget_usd.
| Job | Cap when you give none |
|---|---|
| World build | $20 |
| Re-check of a world | $3 |
| Improvement | $2 |
| Evaluation | $0.06 for each episode in the run |
At a limit
The server refuses the request and says what is left. Nothing is started and nothing is charged.
| Status | Meaning | What to do |
|---|---|---|
402 | The plan does not cover the request: applications, world builds, episodes or the model allowance are used up. | Do not retry. Run fewer episodes, remove an application, add a top-up, or change plan. |
429 | Something is still in progress: a build, or every sandbox of the plan. | Wait for it to finish, or stop it. |
Errors lists the exact messages.
When a payment fails
A workspace keeps its plan while the payment provider retries a failed renewal (its grace period). When the provider reports that the subscription is no longer paid, or the paid period has been over for 3 days with no news of a payment, the limits of the Free plan apply to new work.
Was this page helpful?