Pricing
Pay for what a session actually uses. No monthly subscription fee. During the beta, sign-in with GitHub gives an instant free tier (10 sandbox-hours, 2 concurrent sandboxes, no card); once card verification opens, verifying a card in the console (a $0 check, nothing is charged) upgrades the account to the verified tier (100 sandbox-hours, 10 concurrent). Card verification is not open yet (GET /v1/tiers reports card_verification_enabled), so every self-serve key is on the free tier for now. Details are on the limits page. Billing opens after the beta; the prices below apply then.
Price sheet
| Item | Price |
|---|---|
| CPU | $0.040 per consumed vCPU-hour — measured host-side; waiting on the model or on I/O is not consumed CPU |
| Memory | $0.003 per configured GiB-hour while the sandbox is running |
| Paused sandboxes | No CPU or memory charge |
| Snapshots | $0.10 per GiB-month — snapshots are not yet available and not yet metered |
| Internet egress | First 10 GiB per organization per month included, then $0.02 per GiB — egress is not yet metered during the beta |
| Subscription | None |
Worked example
A typical coding-agent session averages 0.05 consumed vCPU with 4 GiB configured:
0.05 × $0.040 + 4 × $0.003 = $0.002 + $0.012 = $0.014 per running hour
Two fully busy vCPUs with 4 GiB configured:
2 × $0.040 + 4 × $0.003 = $0.080 + $0.012 = $0.092 per running hour
Time spent paused adds nothing to either figure.
GPU pricing
GPU sandboxes are VRAM slices of a shared card, billed per second of real use at the price of the class you pick, in dollars per busy GiB-hour. Every class sells from 0.5 GiB. A class marked "Live now" can be created today. A class marked "On request" cannot be created yet: the price in its row is the list price that applies once it is live, and you can name the class in the use-case field when you request beta access. Idle time, and paused leases on pause-capable templates, bill nothing; templates that cannot release the GPU bill while they hold it.
| Class | Availability | List price per busy GiB-hour | 0.5 GiB slice, per busy hour | 4 GiB slice, per busy hour | Dedicated whole card, per running hour |
|---|---|---|---|---|---|
nvidia.5090 | Live now | $0.13 | $0.065 | $0.52 | $1.55 |
nvidia.rtx6000 | On request | $0.10 | $0.05 | $0.40 | $0.90 |
nvidia.a10 | On request | $0.18 | $0.09 | $0.72 | $1.68 |
nvidia.a6000 | On request | $0.08 | $0.04 | $0.32 | $1.42 |
nvidia.a100-40g | On request | $0.17 | $0.085 | $0.68 | $2.59 |
nvidia.gh200 | On request | $0.08 | $0.04 | $0.32 | $2.98 |
nvidia.h100-pcie | On request | $0.14 | $0.07 | $0.56 | $4.28 |
nvidia.h100-sxm | On request | $0.18 | $0.09 | $0.72 | $5.58 |
nvidia.b200 | On request | $0.13 | $0.065 | $0.52 | $9.09 |
Availability as of 2026-10-18. Machine-readable: gpu-classes.json.
What counts as busy: every exec is billed from dispatch to result (success or failure); idle residency is billed only for intervals where the GPU is observed active. An interval the sampler cannot classify is billed only for a suspend_mode: "none" template, which holds the card exclusively — every other lease bills it nothing, because we do not charge a second we could not classify. Suspended and truly idle time bills 0. Exception: a suspend_mode: "none" template bills its whole residency, because it holds the card exclusively; a freeze-only pause bills its frozen intervals in full for the same reason. Resuming a paused lease also bills the resume time. The reply's meters.billed_s is the number you are billed on.
Idle and paused sandboxes cost $0. Example: a 1 GiB sandbox that is busy 38% of the time it is up (the measured average) costs about 0.38 × its class's price per busy GiB-hour, for each hour it is up.
Whole-card providers bill idle time; here an idle slice costs nothing.
Dedicated whole card: the same card without co-tenants, billed per running second including idle time, availability on request — name the class in the use-case field when you request beta access. The per-class price is in the table's dedicated column.
How we price: each class's list price is the hourly rent of the card we resell, divided by (VRAM × 50% target busy utilization × 60%) — the 60% is what is left of revenue after a 40% target gross margin — and rounded up to the cent; break-even is at about 30% busy. Rent comes from the providers' live pricing pages (Lambda and RunPod), read on 2026-10-18.
Estimate a month
— per month · busy — · idle —
No list price yet for this class. Beta use draws down your GPU allowance.
Idle is $0 on pause-capable templates; a suspend_mode: "none" template (for example mujoco) bills its whole residency as busy. The first GPU-hour per month on a free account and the first 2 GPU-hours per month on a verified account are included; the estimate above is before that allowance.
Beta status
- Free tier: sign in with GitHub, 10 one-time sandbox-hours, 2 concurrent sandboxes, no card. It includes a small GPU trial (see GPU pricing).
- Verified tier (card verified, $0 check): 100 one-time sandbox-hours in total, 10 concurrent sandboxes. Once card verification opens, one card upgrades one account.
- Free accounts: 1 GPU-hour per month included, no card. The allowance refills monthly; 1 concurrent GPU lease, a slice of up to 2 GiB, no internet egress.
- Verified accounts: 2 GPU-hours per month included. The allowance refills monthly; 1 concurrent GPU lease.
- Waitlist-approved beta accounts keep their one-time credit of 100 sandbox-hours. Paused time doesn't consume credit.
- Billing opens after the beta; the prices above apply then.
- First-cohort price lock: 90 days.
Also see: limits · benchmarks