Skip to main content
Replicate lists $0.000100 per second for its CPU hardware configuration. For public models priced by hardware time, it charges for active processing, not instance setup or idle time. Model that charge with a SUM meter for active seconds on CPU instances, rather than a per-prediction fee.
Replicate hardware rate table with per-second and hourly rates, CPU and GPU specifications, and hardware requiring committed spend. The CPU row displays $0.000100 per second.

Replicate's complete hardware pricing table, including hardware offered on committed-spend contracts. Only the CPU row for time-priced public predictions is modeled. Captured November 18, 2026. Source: Replicate's official pricing page, linked below.

View the original pricing page. Scope: Gross active CPU runtime cost for hardware-time-priced public model predictions only. Not modeled: Public models with output-based rates, other hardware, private models, deployments, training, failure and cancellation adjustments, prepaid credit collection, taxes and discounts.

Pricing profile

What Replicate charges

Design takeaways

  1. Choose the interval that drives the charge. Client-observed elapsed time can include nonbillable work. Carry provider-equivalent active seconds, not request latency, to the meter. See the published terms.
  2. Filter the pricing class separately from the hardware. A public prediction is not automatically billed by runtime. Keep output-priced predictions out of the CPU seconds meter even when the event stream covers all predictions. See the published terms.
  3. Keep prepaid funding out of the usage formula. Replicate describes credit balances for prepaid accounts. The Flexprice example calculates gross usage before applying any funding or payment collection. See the published terms.

How Replicate’s pricing maps to Flexprice

Configure the metered subscription

These examples are illustrative. Replace example IDs with IDs returned by your Flexprice environment. Create a metered feature with POST /features:
Create the plan with POST /plans:
Create the usage price with POST /prices. Use the returned meter ID, not the feature ID. The USD price bills at the end of each period:
The usage price is a constant rate per second, not a recurring fee. There is no fixed subscription price, volume tier or quantity transformation. Create one customer for the account. If you configure a metered entitlement, make it unlimited so it does not cap the priced usage. Start the subscription with POST /subscriptions:

Send finalized usage

Set external_customer_id to the external ID of the same customer used on the subscription. The event uses integration-defined properties, not a provider response:
Supply finalized active billable seconds for predictions known to be on the CPU hardware rate. Preserve precision; reconcile your input with priced usage. Exclude setup and idle time, rather than subtracting an assumed startup time from request latency. The monthly arrears schedule here is illustrative, not a claim about Replicate’s collection window. Keep the same event ID and original timestamp when retrying. Do not emit both individual records and an interval total covering the same usage. Aggregate before rounding currency.

Worked bill

133,750 active CPU seconds × $0.000100 = $13.375 before currency rounding. An individual prediction with 12.375 rated seconds contributes $0.0012375 to that total before currency rounding. Do not round individual predictions to cents.

Pricing sources

Pricing checked on 2026-11-18. Recheck the source before implementing. This is illustrative configuration, not a recreation of a full provider invoice. Usage-based pricing recipe