Skip to main content
Docs

Search guides and API endpoints, for example “Idempotency-Key” or “submit job”.

    Concepts

    Wallet and billing

    What is charged to your wallet, how reservations work, and the price per hour of every machine on every plan.

    Updated October 1, 2026

    On this page

    Compute is paid from a prepaid wallet in US dollars. Your plan sets the per-second rates; the wallet pays for what you run.

    Funding the wallet

    • Deposits: add funds by card on your billing page (payments go through Stripe).
    • Monthly credit: paid plans add credit to the wallet each time the subscription renews; see Plans and limits.
    • Vouchers: redeem a voucher code from your account.

    Your account lists every deposit, credit, and charge.

    What is charged

    Charged to the walletNot charged
    Jobs: seconds run × your plan's rate for that hardwareEstimates and workflow validation
    Workflow runs: the sum of their steps' jobs (a step reused from an identical earlier run costs nothing)Reading status, results, artifacts, and events
    Assistant turns: the tokens each turn usesScience lookups and reference papers
    Utilities and inference: they count toward monthly allowances instead

    Reservations

    Starting work first reserves an amount, so you can never start more than you can pay for:

    • A job reserves its estimate when you submit it.
    • A workflow run holds its estimate when it starts; each step draws on that hold as it runs.
    • An Assistant turn holds enough for the longest reply it could give.

    When the work ends you pay the actual cost and the rest of the reservation is released. Get wallet balance returns your balance and active_reserved; what you can still start is the difference. If a new job or run doesn't fit, the request fails (400 for jobs and workflow runs, 402 for Assistant turns and plan-card approvals) and nothing starts.

    Spend ceilings on API keys

    An API key can carry a spend ceiling: the most that work started with that key may reserve in a calendar month (UTC). A job or run that would pass it fails with 403 (spend-ceiling-exceeded); the count starts again each month. Set one on every key you give to a script or an AI client; see Authentication.

    Price per hour

    What one hour of each machine costs on each plan. Jobs are charged per second.

    Compute price per hour, USD
    HardwareBasicStarterProEnterprise
    CPU$0.50$0.29$0.18$0.11
    Nvidia T4$2.05$1.48$1.19$0.90
    Nvidia L4$2.81$2.02$1.58$1.19
    Nvidia A10$3.85$2.74$2.20$1.66
    Nvidia L40S$6.84$4.90$3.89$2.92
    Nvidia A100, 40 GB$7.34$5.26$4.21$3.17
    Nvidia A100, 80 GB$8.75$6.26$5.00$3.74
    Nvidia H100$13.82$9.86$7.88$5.94
    Nvidia H200$15.88$11.34$9.07$6.80
    Nvidia B200$21.89$15.62$12.49$9.40

    Assistant token prices are in The Assistant.