Rate card

Model

Price is the sum of four components:

total  =  tier time rate × seconds in running
       +  resource meters (cpu, memory, disk, gpu)
       +  network (egress, ingress, transfer)
       +  storage (volumes, snapshots, recordings)

The tier time rate is the large term. It pays for the isolation boundary, the warm pool that holds the start-latency promise, and the capacity reserved at that tier's oversubscription ratio. Resource meters are a smaller correction for what the session actually consumed.

This is why tier selection dominates a bill far more than optimising the agent's code.

Tier time

Rates are for the m size class (2 vCPU, 4 GiB). Larger classes scale roughly linearly in vCPU and memory.

Resources

Network

Storage

paused_gib_month is five times the volume rate because paused memory occupies real RAM on a real host that cannot be used for anything else. Pause is for seconds to minutes; snapshot is for longer.

Reserved capacity

Reserving a tier 4 warm instance costs 60% of a running tier 4 session, continuously. A guarantee of twenty concurrent sub-10-second desktop starts is twelve desktops' worth of continuous spend before anyone runs anything.

Worked examples

A support agent, tier 0 dominant

10,000 conversations per day. Each: 3 tool calls, one 4-second tier 1 code run.

tool_call      30,000/day × $0.00001                        = $0.30 / day
tier1.session_s 10,000 × 4 s × $0.00002                     = $0.80 / day
cpu_ms          10,000 × ~3,000 ms × $0.00000002            = $0.60 / day
                                                     total    $1.70 / day   ≈ $51 / month

A coding agent, tier 2 dominant

500 sessions per day, 8 minutes each, 2 vCPU / 4 GiB, 200 MiB egress each.

tier2.session_s 500 × 480 s × $0.000062                     = $14.88 / day
cpu_ms          500 × ~240,000 ms × $0.00000002             = $ 2.40 / day
mem_gib_ms      500 × 4 GiB × 480,000 ms × $0.0000000025    = $ 2.40 / day
egress_gb       500 × 0.2 GiB × $0.010                      = $ 1.00 / day
                                                     total    $20.68 / day  ≈ $620 / month

A browser agent, tier 3

200 tasks per day, 90 seconds each, 3 pages open on average, 50 MiB egress.

tier3.session_s  200 × 90 s × $0.000112                     = $ 2.02 / day
browser_page_s   200 × 90 s × 3 × $0.000004                 = $ 0.22 / day
egress_gb        200 × 0.05 GiB × $0.010                    = $ 0.10 / day
                                                     total    $ 2.34 / day  ≈ $70 / month

A computer-use agent, tier 4

50 tasks per day, 6 minutes each, plus 2 reserved warm instances.

tier4.session_s     50 × 360 s × $0.000375                  = $ 6.75 / day
pool_reservation_s  2 × 86,400 s × $0.000225                = $38.88 / day
                                                     total    $45.63 / day  ≈ $1,369 / month

The reservation is six times the usage. That is the honest price of a latency guarantee at tier 4, and it is the number that decides whether a customer wants one.

What drives a bill

In descending order of impact:

1. Tier selection. A workload moved from tier 3 to tier 1 is roughly 6× cheaper per second.
2. Session lifetime. Sessions left to idle out, rather than deleted, pay the whole timeout.
3. Pool reservations at tier 4. Often larger than usage itself.
4. Egress, for scraping and data-moving workloads.
5. Idle volumes. Provisioned storage nobody is using.
6. Open browser pages. Multiplied by page count, easy to miss.
7. Resource meters. Usually the smallest term, and the one teams optimise first.

Discounts

Committed use

A contractual monthly minimum in exchange for a discounted price list, applied to every meter.

Terms are 12 or 36 months. Usage below the commitment is still charged at the commitment — the discount buys a rate, not an option.

Volume tiers

Applied automatically to the largest meters, no contract required.

Marginal, not retroactive: the first $2,000 is always at list.

Preemptible

A 55% discount in exchange for eviction with 30 seconds' notice. Available at tiers 1–3. Correct for batch evaluation, bulk scraping and anything that can checkpoint and resume; wrong for interactive agent work.

Free tier

Every organization, every month, no card required:

50,000  tool calls
 2,000  tier 1 session-seconds
   500  tier 2 session-seconds
    10  GiB egress
     5  GiB-month storage

Enough to build and test an integration end to end. Tiers 3 and 4 are not in the free tier because one session of either would consume it.

Billing mechanics

Rating is per minute bucket and summed, not rounded per session. A session running 90 seconds bills 90 seconds, not two minutes.