Prepaid credits, drawn down per second at the rates below. No multipliers,
no preemptible tier, no eviction: a sandbox is yours until you release it.
Live via GET /v1/pricing.
| price | |
|---|---|
| vCPU-hour, 2 GiB of memory included | $0.06 |
| additional memory, per GiB-hour | $0.02 |
| minimum sandbox | 1 vCPU, 2 GiB ($0.06/hr) |
| standard agent trial (2 vCPU, 4 GiB) | $0.12/hr |
Pick any shape. Sizes double from 1 vCPU / 2 GiB to 64 vCPU / 128 GiB;
shapes between sizes bill as vCPU plus memory at the rates above. Memory is
never oversubscribed. Seconds lost to a fault on our side (provider_*
causes) are never billed.
Docker-in-Docker, Compose and Kubernetes sandboxes bill at the same rates.
| price | |
|---|---|
| continuous checkpoints (memory and disk at every tool call) | included |
| retained checkpoint storage, per GiB-month | $0.08 |
| suspended sandbox memory, per GiB-month | $0.12 |
| persistent volumes, per GiB-month | $0.10 |
| export to your own in-region S3 bucket | free |
| artifact egress elsewhere, per GiB | $0.02 |
Retention has two clocks, both starting when a sandbox ends. The memory
clock (default 24 hours) drops the memory bases and diffs; every checkpoint
remains as a filesystem checkpoint you can browse, diff and fork with fresh
processes. The full clock (default 30 days) retires the checkpoint. Both are
yours to set per sandbox or per org, in seconds or by preset (default,
memory_7d, 7d, 30d, forever); 0 means never. Pin a checkpoint or a
trial and it is exempt from both clocks at the standard rate.
An example, typical of an agent benchmark: 60 trials, about an hour each, 2 to 4 vCPU, a few hundred tool calls per trial. Compute for the run is about $26. The run leaves about 2,400 GiB of checkpoints, and about 78 percent of that is memory (bases and diffs); the rest is the filesystem and metadata. At $0.08 per GiB-month:
| you keep | memory | filesystem | total storage cost for the run |
|---|---|---|---|
default | 24 hours | 30 days | about $45, once (about $5 of memory for a day, about $40 of filesystem for the month) |
| memory 24 hours, filesystem 7 days | 24 hours | 7 days | about $12, once |
memory_7d | 7 days | 30 days | about $75, once |
7d | 7 days | 7 days | about $44, once |
30d | 30 days | 30 days | about $188, once |
forever | until you delete | until you delete | about $188 every month you keep it |
| pin one trial's checkpoints, default for the rest | the default, plus about $3 a month for as long as the pin stands |
Every row except forever is a one-time cost: when the window ends the
checkpoints are gone and so is the charge. Storage is metered per
GiB-month and billed per second like everything else, so a window shorter
than a month costs that fraction of a month.
Memory is what a fork or a rewind with processes intact needs, and that is
done while a run is being looked at, so a day of memory covers it. The
filesystem is what you browse, diff and export for weeks, and it is a
fifth of the bytes. That is why the default is shaped the way it is, and
why anything longer is a choice you make with the number in front of you:
storage is its own line on every run and trial, never inside compute.
Storage is billed and shown as its own line (storage_usd, retained_gib
on every run and trial), never inside a trial's compute cost.
| price | |
|---|---|
| H100 80 GB | $3.89/hr |
| RTX PRO 6000 96 GB | $2.99/hr |
| CPU and memory attached to a GPU sandbox | compute rates above |
Billed per second, never preemptible. A GPU sandbox is yours until you release it.
A gpu_fraction of a card bills that fraction of the card rate by
occupancy: a quarter of an H100 is a quarter of its compute and memory for
$0.97/hr. Temporal sharing (several sandboxes taking turns on one card)
bills each sandbox for the seconds it holds the card. Oversubscribed
fractions (gpu_oversubscribe=true) bill at 0.6 of the guaranteed
fraction's rate; guaranteed fractions keep their seats.
A group of GPU jobs with a deadline pays a fraction of the card rate set by the slack at submit and fixed for the batch:
| slack at submit | fraction of the card rate |
|---|---|
| under 1 hour | 1.00 |
| 1 to 6 hours | 0.85 |
| 6 to 24 hours | 0.70 |
| a day or more | 0.60 |
GET /v1/gpu/batches/quote?deadline=24h returns the exact rate before you
submit. At worst a job pays the normal card rate.
Monthly and annual commitments are available for teams above $10k a month; talk to us at hi@numinous.technology.