Pay for what runs
Prepaid credits and no subscription. A run is charged for the seconds its worker is alive, the model calls it makes and the web requests it sends, and never more than the budget you give it.
How a run is charged
1Reserve
Before the worker starts, the run's budget is set aside from your available credits.
2Run
Compute accrues per second; each model call and web request is authorized against the budget, then charged.
3Settle
When the worker is destroyed, the run is charged for what it used and the rest is released.
Estimate a run
Use the prices of the model you plan to call. The dashboard shows each model's price when you pick it.
Measured runs
Real runs of the recipes in the docs on the hosted service, against the sample API, with openai/gpt-6-luna, the default model. Each is the median of three runs measured on Oct 7, 2026, priced at today's rates.
- Investigate a declined payment$0.002618Run time 12 s · Model calls 3 · Tokens 5,101 · Web requests 0
- Investigate a payment that failed during an incident$0.004297Run time 15 s · Model calls 4 · Tokens 7,261 · Web requests 1
- Check a status page$0.005014Run time 19 s · Model calls 2 · Tokens 2,445 · Web requests 1
- Write the weekly report$0.002245Run time 10 s · Model calls 2 · Tokens 2,486 · Web requests 0
Self-hosted
Run Ephemeral in your own AWS account with your own model provider. Licensed per installation under an enterprise agreement; infrastructure and model costs are paid to those providers directly.
Read the self-hosting guide- Same engine and dashboard
- Your VPC, keys and model endpoint
- Limits on installations, concurrency and monthly runs
Questions
Do I pay anything while I'm not running agents?
No. There is no subscription or minimum. You buy credits when you need them, and only runs spend them.
How much will a run cost?
Mostly whatever the model charges. Compute is $0.0002 a second, so a 30-second run adds $0.006. Model calls cost what the provider charges for them. Try the estimator above; every run's page shows its exact cost.
Could a run ever cost more than I expect?
No. Every run has a budget: $0.50 unless you set one, and up to $100. A model call that would go over it is refused, and the worker is stopped at its timeout.
Do I pay for a run that fails?
Only for what it used before it stopped: its seconds of compute and the model calls it made. The rest of its budget goes back to your balance. If a call is cut off before the provider reports its cost, that call's authorized amount may be charged.
What happens when my credits run out?
New runs don't start until you add credits; runs that already started keep the budget they reserved. Automatic top-up can add credits for you when the balance gets low.
Do you mark up model prices?
No, not on models. Each call costs what the provider reports for it, and you pay us for compute and web requests. The one addition is zero data retention: a run that uses it costs 10% more, on compute, model calls and web requests.
What do web requests cost?
$0.001 for each page or API request and $0.01 for each search. They come out of the run's budget like model calls, and reading further into a page the run already fetched is free. Web access is off unless you turn it on for a run.
Do schedules cost anything?
Only the runs they start. A scheduled start is an ordinary run that reserves its budget and is charged for what it uses; the schedule itself is free.
How do I add credits?
By card through Stripe, from $5 to $10,000 at a time. Credits appear once Stripe confirms the payment, usually within seconds. You can save the card and turn on automatic top-ups.
Do credits expire? Can I get a refund?
Credits never expire. Within 14 days of a purchase, you can ask through support for the credits from it that you have not used to be refunded to your card. The Terms of Service have the details.
Can my team share credits?
Yes. Credits belong to an organization, and every member's runs draw from the same balance. Owners and admins handle payments.