Two cents a sandbox-minute
You pay for the time your trials hold a sandbox, by the second, and only for trials that were scored. Your model’s calls stay on your own key.
$20,000
in credits for every new workspace, good for 90 days.
About 34 full Terminal-Bench 4.1 runs at our planning figure of 40 minutes a trial.

Know the price before you launch
A benchmark’s official rules fix its tasks and trials. What is left is how long a trial takes, and that is the one thing you are charged for.
Worked here for Terminal-Bench 4.1 by its leaderboard’s rules, at our planning figure of 40 minutes a trial. With scripts on, the figures can be changed.
Estimate$0.02 a sandbox-minute
Terminal-Bench 4.1
66 tasks, 5 trials each
- Trials
- 330
- Sandbox time
- 13,200 min
- Sandbox size
- 2.2 standard
- Sandbox-minutes
- 29,040
- Wait, 8 at a time
- 27h 30m
$581
$580.80 from your credits if every trial is scored. The $20,000 a workspace starts with covers it 34 times. Your model’s calls are billed by your provider, to your own key.
One thing is charged
What your credits pay for is the sandbox time on your run pages, added up and multiplied by the price. Nothing else draws on them.
Charged
The time a scored trial holds a sandbox
From the moment its try starts to the moment it is graded, by the second. A sandbox-minute is one minute of a standard sandbox, 2 vCPU and 4 GiB. A benchmark whose tasks ask for larger sandboxes is charged in proportion, and its rate is shown before you launch.
Not charged
A trial that ended in an error
A sandbox that would not start and a provider that was down are not yours to pay for. The trial is left out of the score and costs you nothing.
Not charged
Waiting
A trial in the queue holds no sandbox, and a run that is paused or halted by its budget has no clock.
Not charged
Your model’s calls
Your provider bills them to your own key. Berkway records what each one cost and never handles that money.
Not charged
Berkway Smoke
Three small tasks that prove your endpoint, the sandbox and the grading work. Run it as often as you like.
Three ways to buy it
Every plan is the whole product, at the same price a minute. What changes is how you buy credits and how closely we work with you.
Run
$0.02a sandbox-minute
Pay as you go. Credits are spent at face value, and every new workspace starts with $20,000 of them.
- Every benchmark, every agent and every setting
- Traces, budgets, intervals and error accounting on every run
- A record of every run, yours to publish
- Your own tasks, private to your workspace
Lab
$1,500a month
For a team that runs every week. $1,500 a month buys $2,000 of credits, every month.
- Everything in Run
- Your own benchmarks set up with you, scorer included
- A shared channel with the engineers who run the service
- Each new benchmark run on your models the week it is added
Contract
By agreement
For a company whose purchasing needs more than a card and an invoice.
- A price for your volume
- A data processing agreement and a security review
- Purchase orders and your payment terms
- A named person for your account
Prices are in US dollars, exclude taxes, and are as decided on November 21, 2026. During the pilot no payment is taken: credits are granted and plans are arranged with a person. Card payments open later, and nothing is charged before they do.
Start with $20,000 in credits
Berkway is in pilot, and access is by request. Credits are good for 90 days. When they run out, runs stop until more are added, so nothing is ever spent that you did not see.
1
Ask
Tell us which model you want to measure. A person reads every request and answers within a working day.
2
Connect
Create a workspace with the code we send and connect an endpoint: any OpenAI-compatible base URL, with your own key.
3
Prove it
Run Berkway Smoke. Three small tasks show that your endpoint, the sandboxes and the grading work, at no charge.
4
Run
Launch a benchmark. It draws on your $20,000 at the list price, and the form shows how much before you launch.
Questions about price
How do credits work?
Every new workspace starts with $20,000 in credits, good for 90 days. A run draws on them at the list price, for the sandbox time of its scored trials. When they run out, runs stop until credits are added. Credits are not money and cannot be withdrawn, transferred or refunded.
What is a sandbox-minute?
One minute of one trial holding a standard sandbox, 2 vCPU and 4 GiB. Thirty-two trials running for a minute are thirty-two sandbox-minutes. The clock is per trial and counts seconds. A benchmark that needs larger machines counts each minute as more than one, and the launch form shows its rate.
How do I know what a run used?
Every run page shows its sandbox time, counted over scored trials only, and what it drew from your credits. The usage page of your workspace adds a month up run by run, so you can check every dollar against the runs that spent it.
Does running more trials at once cost more?
No. Concurrency changes how long you wait, not how many sandbox-minutes the trials use. A benchmark costs the same at four at a time as at sixty-four.
Can I cap what a run costs?
A run’s budget is set in dollars of model spend and halts the run at the call that reaches it. Trials that were still working go back to the queue unscored, so they draw nothing. Each task also keeps its benchmark’s own time limit.
Do you add anything to my model bill?
No. Calls go to your endpoint with your key, and your provider bills you. Berkway takes no share and adds no margin.
What if a trial is tried again?
A trial is tried again only after a failure that was not the model’s. You pay for the try that was scored, not for the ones before it.
How do I pay?
During the pilot no payment is taken. A person grants your credits and arranges a Lab or Contract plan with you. Card payments open later, and nothing is charged before they do.
Will the price change?
It may, with thirty days’ notice by email, and never for a run already launched. A Lab or Contract price holds for its term.