Skip to main content

Execution Time Pricing

Alongside token charges, every node that runs accrues a small charge for the time it spent running. It's the component people don't expect, and the one that explains why two flows with identical model usage can cost different amounts.

Why running time costs money​

Token prices pay the AI provider. Execution time pays for the infrastructure your flow actually runs on — the machines holding your flow in memory, waiting on APIs, and doing everything between the model calls.

A node that waits nine seconds for a slow third-party API used no tokens at all, but it occupied a worker for nine seconds. That's what this charge covers.

How execution time is measured​

Per node, in wall-clock seconds, from when the node starts to when it finishes. Summed across every node in the run.

Wall clock is the important part. Time spent waiting counts: a slow HTTP call, a large document being processed, a model taking its time. The flow isn't doing anything during the wait, but the capacity is still reserved.

Per-minute rate​

The rate appears in the Exec (per min) column of the Pricing tab — at the time of writing, $0.12 per minute, or $0.002 per second.

Some arithmetic to calibrate:

FlowTotal execution timeExecution charge
Start → Agent → Direct Reply~3 seconds~$0.006
Retriever + Agent + two tools~10 seconds~$0.02
Multi-agent flow with an external API~30 seconds~$0.06

Small per message. Meaningful at ten thousand messages, and meaningful in a batch run that fires a thousand of them at once.

What makes a flow run longer​

In roughly descending order of impact:

  • Slow external calls. An HTTP node or tool that waits on a third party is usually the largest single contributor. Nothing about the flow is inefficient — the other end is slow.
  • Long model responses. Generating 800 words takes measurably longer than generating 80.
  • Node count. Every node adds its own time, however small. A twelve-node flow that a six-node flow could do costs twice the execution time.
  • Large retrievals. Fetching and embedding more passages takes longer than fetching fewer.
  • Iteration and loops. Iteration and Loop multiply everything inside them.
note

A Wait node is billed differently. The wait suspends the run rather than occupying a worker, so it doesn't accrue execution time — but the duration itself is charged, at its own much lower per-second rate. A full-day wait costs a few cents, not the dollars the same span of execution time would. A single Wait node can't exceed 24:59:59.

Finding the expensive flows​

The Usage Dashboard's distribution view breaks cost down by flow and by node. When a flow costs more than it looks like it should, the node-level breakdown usually shows one step accounting for most of it — and it's usually an external call, not the model.