Skip to main content
Compute is drawn from the organization’s real-USD balance: a computer you create is billed ahead in 10-minute blocks, and a session’s computer is metered from the seconds it ran, booked against your budget at the end of each turn. This page is the reference for the caps and the exact per-unit rates.
Every rate on this page is the price you pay — the number booked to the ledger and reported on your spend breakdown. Model inference is priced separately; see Model router & inference.

The compute meter

A computer is priced on its provisioned resources — vCPU and memory — at the rates below. There is no creation fee, and disk is not metered: the boot disk is inside the allowance the computer ships with. A computer you create yourself (create) is billed ahead, in 10-minute blocks of the clock (:00, :10, :20, …):
  • At create, it is charged from its creation to the end of the current block. A computer that runs for 30 seconds pays for up to ten minutes.
  • While it runs, each later block it runs in is charged in full. A computer paused mid-block keeps the rest of that block; a resume inside a block already paid is not charged again.
  • Create and resume need a balance that covers one full block at the computer’s size; otherwise they answer 402 insufficient_credits. Running a command or a file operation on a paused computer resumes it, so it needs the same.
  • When a block cannot be paid in full, or the organization’s plan has lapsed, the computer is paused, not destroyed. Its disk is kept, and it can be resumed once the balance covers a block.
A session’s computer is metered per running second instead, from the time its turns actually used it. A session’s computer is 1 vCPU / 1 GiB, so a full hour of running time is $0.0504 + $0.0162 = $0.0666. Because idle maps onto the runtime’s wake/sleep cycle, a computer waiting between turns is metered for nothing at all — the mechanism behind Naive Managed Agents’ low cost on multi-hour tasks.

The browser

The optional browser is metered per session minute, open to close (rounded up, minimum one minute), as one browser line on the ledger debited when the session closes — there is no up-front hold, because a session’s cost is knowable only at close. The model-driven actions (act, extract with an instruction, observe) add their spend to the same browser line, one entry per action. A browsing agent also pays the tokens it spends driving the browser and the computer time of the session driving it. Its limit is the per-session timeout below; a session past it is closed and billed by the platform clock within a minute.

Mobile devices

A mobile device is metered per phone-minute, create to end (rounded up, minimum one minute), at $0.05 a minute for an Android phone and $0.08 for an iPhone, each including the residential network its traffic exits through. It is one mobile line on the ledger, debited once when the device ends, with no up-front hold. Actions, app installs included, add nothing. A device that failed to provision bills nothing. A device stopped at its max_duration_minutes is billed for the minutes it actually ran until it ended, at most 5 minutes past its maximum. There is no limit on how many devices an organization runs, only on what it can pay for: a create needs a balance that covers 10 minutes for every live device and the new one, each at its rate ($0.50 for one Android phone, $0.80 for one iPhone), plus what the live ones have run so far (402 insufficient_credits otherwise). The platform checks running devices every minute: an organization whose plan has lapsed, or whose balance no longer covers what its devices have accrued, has them ended with termination_reason: "unfunded", billed for the minutes they ran. A past_due plan keeps its devices through the payment grace period.

Concurrency

There is no count limit on computers or mobile devices. What an organization runs is bounded by what it can pay for.

Boot & wake latency

Only cold boots pay the full provisioning cost; the wake/sleep cycle the runtime uses between turns is near-instant, which is what makes idle cheap.
The fully virtualized isolate mode (mode: "isolate", coming soon) targets near-instant boot and near-zero idle; the numbers above describe the micro-VM that ships today.

Limits

The 12-hour max runtime bounds a computer’s total lifetime from create to destroy — it is wall-clock, not billed time, so a computer that is paused for most of that window still hits the ceiling on schedule. Only running time is billed; the ceiling is a safety cap on how long the computer may exist.
A computer is auto-destroyed when it exceeds its max runtime. Persist anything you need to keep with publish_file or a checkpoint before then.

Inspecting spend

Compute spend appears as the computer component of the agent’s spend breakdown and as a line item on the organization ledger.
CLI
Spend is reported on the wire as integer micro-USD (1 USD = 1_000_000); the CLI renders dollars for humans.

Back to the computer overview

The full computer surface at a glance.