# Pricing > How Nodus sets your rate, what each second and token costs, and how to see the price before you launch. Source: https://nodus-platform-site.pages.dev/docs/concepts/pricing/ Build revision: 211ad9f836655b1c3a2668c4693e442471f28614 Nodus is prepaid: you buy credits, every run shows its rate before it starts, and the rate stays frozen for that run. The numbers on this page come from the same price list the API uses, so the [pricing reference](https://nodus-platform-site.pages.dev/docs/reference/pricing/) and your estimates always agree. This page explains how those numbers are set. ## See the price before you launch `nodus run --dry-run` and the console review screen show the estimate for a run: the hourly rate, the expected cost range, the expected boot and teardown time as their own lines, and the minimum charge. The estimate is valid for up to 31 minutes. Launching with the estimate’s `ETag` holds you to it: if prices change in between, the launch is refused and you review the new estimate. ## Machines Nodus rents for you A Job, a GPU Sandbox, a GPU Workspace or a GPU Function worker runs on a machine Nodus rents for you. For that machine you pay exactly what the provider bills, divided by 0.875, so Nodus keeps 12.5 % of what you pay: * **From creation to deletion.** The clock runs from the moment the provider starts billing until the machine is confirmed deleted. Your usage is itemized by segment: `Boot` (start-up, image pull, readiness), `Restore` (resuming from a checkpoint), `Running` and `Teardown` (stop to deletion, plus the provider’s billing increment, charged once per machine). * **Never above list.** Every accelerator and count has a list price. Nodus never places your work on a machine whose rate would exceed it. * **Frozen for the run.** The rate is fixed when the machine is acquired. A later price change applies only to runs launched after it takes effect. Spare machines Nodus starts to finish your work sooner, and failures Nodus causes, are never charged to you. The list price is the ceiling; the “from” rate is the lowest rate a machine is available at right now, refreshed every minute. Your estimate shows the rate for your run. |Accelerator|×1 list|×1 from|×8 list|×8 from| |-|-|-|-|-| |A10|$0.89|List only|$7.12|List only| |A100 40G|$1.49|List only|$11.92|List only| |A100 40G PCIE|$1.39|List only|$11.12|List only| |A100 80G|$1.99|List only|$15.92|List only| |A100 80G PCIE|$1.89|List only|$15.12|List only| |B200|$6.49|List only|$51.92|List only| |H100 PCIE|$2.99|List only|$23.92|List only| |H100 SXM|$3.29|List only|$26.32|List only| |H200|$4.29|List only|$34.32|List only| |L4|$0.89|List only|$7.12|List only| |L40S|$1.29|List only|$10.32|List only| |RTX 3090|$0.59|List only|$4.72|List only| |RTX 4090|$0.69|List only|$5.52|List only| |RTX 6000 ADA|$1.09|List only|$8.72|List only| |RTX A6000|$0.79|List only|$6.32|List only| USD per hour for the whole machine. Pricebook 2026.10.8, effective 2026-10-01. ## Sandboxes, Functions and builds CPU Sandboxes, Functions, agent workers, image builds and CPU Jobs placed on Nodus nodes are billed per second from placement to release, at published rates per vCPU, per GiB of memory and per GiB of disk above 10 GiB per vCPU. The smallest billable shape is 0.25 vCPU with 512 MiB of memory; smaller requests are billed at that shape. ## Models Each model’s price is the model’s cost divided by 0.95. When a model is served from more than one source, it is listed once, at the price of the cheapest source available now, and a request is charged the price it was accepted at even if another source finishes it. A request is priced once: every input, cached, output, audio or speech unit is added up exactly, then rounded up to the next micro-dollar. A model is served only for the operations it has a price for. If Nodus cannot confirm the outcome of a request, you are not charged for it. Indra (`nodus/indra`) picks one of 10 models for each request. On the Indra plan you pay the chosen model’s rates plus the routing call (0.044211 USD per 1M input tokens), in one price per request. Without the plan, free Indra routes among the three cheapest models at no charge, up to 100 requests and 200,000 tokens per organization per UTC day. The Indra plan costs 20.00 USD a month and includes 20.00 USD of `nodus/indra` usage each month at the listed prices. The allowance expires at the end of each month, and usage beyond it draws from your credits. Agents that run on Nodus-managed models are billed from your credits at the model’s cost divided by 0.875, on the agent run. The plan allowance does not apply to them. ## Storage, egress and credits * **Storage:** 10 GB per org is included; beyond it, retained bytes are billed per GB-month. * **Egress:** 10 GiB per org per day is included; beyond it, egress is billed per GiB up to your daily egress quota. * **New organizations** receive 30.00 USD of credit, valid for 30 days. ## Rounding Money is counted in micro-dollars. A running machine is charged as it goes, rounded down, and settled when it stops, rounded up once, so charging in many small windows costs the same as charging once. Current rates for every line are in the [pricing reference](https://nodus-platform-site.pages.dev/docs/reference/pricing/).