Pricing

Pay for what the agents use.

One $20/mo plan per organization, $20 of credit included. Everything past that is real dollars, metered per call — no per-seat fee, and nothing at all for a machine that is sitting still.

Pro

$20/mo

One plan per organization, and it comes with $20 of credit every month. Past that you are billed for what your agents consume and nothing else.

Start building

What that means

  • $20 of credit included each month
  • Pay as you go for more credit — top up the prepaid balance any time
  • No per-seat fee — one plan covers the whole organization
  • Usage metered per tier, booked in integer micro-USD
  • MCP, the SDK and the CLI are not metered

Custom

Talk to us

Committed volume, procurement, or anything the meter does not cover.

Book a call

What that covers

  • Committed-volume pricing
  • Invoicing and PO terms
  • A monthly allowance arranged against your account
  • Security review and DPA

What a bill is made of

Four components, each on your spend breakdown under its own name. Two are quoted before they run and two are metered from what they used.

  • Model inference

    Per token, five tiers

    Input, cached reads, cache writes, output and reasoning are metered as five disjoint tiers, at the tariff for the completion window the request asked for.

  • Computer

    Per second, while running

    vCPU and memory on what the machine was given. Nothing to create one, nothing while it sleeps, and the boot disk is inside the allowance — so a computer waiting between turns meters nothing at all.

  • Web tools

    Per call

    Priced from what the provider charged, so a web call is booked after it runs. A fetch your domain policy refuses is stopped before the request and costs nothing.

  • Media generation

    What the job cost

    Billed once per finished job, at its real cost — a render cannot be quoted before it runs, so it is admitted against the headroom under the agent's per-task cap instead. A job that fails is not billed.

The rates

A session’s computer is 1 vCPU / 1 GiB, so a full hour of it running is $0.0504 + $0.0162 = $0.0666 — and an agent waiting between turns is metered for nothing.

Computer · vCPU$0.0504/hr
$0.000014 per second, while running
Computer · memory$0.0162/GiB-hr
$0.0000045 per GiB-second
Computer · paused$0
no vCPU, no memory, no storage
Computer · created$0
there is no creation fee
web_search$0.002158/call
one query, up to 10 results
web_fetch$0.001079/call
one page read as text
MCP, SDK and CLI$0
the interfaces are not metered
Seats$0
no per-seat fee, however many people use the organization

Every rate here is the price you pay — the number the ledger books and the number your spend breakdown reports.

Full price list

The completion window is the biggest lever

The same model at a different window is a different price, and you choose it per request — pay the interactive tariff only when a person is waiting. Overnight work runs on loose for materially less.

  • immediateAnswers nowHighest tariff
  • priorityAnswers soonLower tariff
  • looseAnswers eventuallyLowest tariff

Measured, not modelled.

Every figure is corrected against vendor invoice brackets — list-price rate cards overstate real cost by up to 4.2×. The method, and the cost per completed task, are on the bench.