What does one ComfyUI image cost on a cloud GPU?
A single image cost us about $0.024 in usage on a rented L4 GPU, and that figure came from the platform's own billing after it settled, not from our own arithmetic. The graphics card is not the whole bill: the processor, the memory and the time the machine sits idle are metered separately. Below are the list prices the main cloud GPU platforms publish today, next to the one image cost we have actually measured and where that number came from.
List prices were read from each vendor's own pricing page on 2026-10-05 and are labelled as such. They change. The measured figure is ours and is the only one on this page that is not a list price.
What did one image actually cost us?
About $0.024 in usage, on Modal, reported by the platform's billing after it settled.
Measured once, on a Modal L4, 2026-10-03. Usage on the Modal account, reported after billing settled: L4 about $0.0173, CPU about $0.0056, memory about $0.0012. Cached rerun, not a cold start. This is a usage figure, not an amount charged to a card.
This is a usage figure on the account, not an amount charged to a card. On that workspace the amount charged for the month was $0.00, because credits covered it. Keep the two apart: they come from different pages and should never be added together. The full table is on the episode 2 page.
What do the platforms list per hour?
List prices read from each vendor's own page on 2026-10-05, for a small GPU that will run a diffusion model. The right-hand column is the only place where a real cost appears, and it is real for exactly one of these four.
| Platform | GPU | List price | Card needed? | Free credit | Our measured cost, one image | Notes |
|---|---|---|---|---|---|---|
| Modal modal.com/pricing | Nvidia L4 | $0.000222 per second list price, read on 2026-10-05 | Yes Modal's docs: a payment method on file is required to use Modal at all. | $30 of credits per month on Starter | $0.024 (see below) | Per-second billing. CPU, memory and storage are metered separately. |
| Runpod www.runpod.io/pricing | L4 | $0.49 per hour list price, read on 2026-10-05 | Not verified The pricing page does not state a card requirement. | Not verified | not tested | Community Cloud, Pods. Their page was last updated 2026-09-27. |
| Vast.ai vast.ai/pricing | L4 | from $0.25 per hour, median $0.32 per hour list price, read on 2026-10-05 | Not verified The pricing page does not state a card requirement. | Not verified | not tested | A marketplace, so the price moves with supply. They publish a low and a median, and both are shown here. |
| Google Colab colab.research.google.com/signup | Not read | Not verified pricing page is a web app; we could not read a price from it | Not verified We did not read a signup requirement. | Not verified | not tested | Listed because people ask about it. We have no verified figure for it, so nothing is shown. |
Two of these are billed per second (Modal and Vast.ai), so a job that runs for 90 seconds is billed for 90 seconds. Read the notes column before you compare the headline numbers.
Do I need to put a card on file?
On Modal, yes. The other three platforms did not say, so those cells read "Not verified" rather than a guess.
Modal's own billing documentation is explicit: you must have a payment method on file in order to use Modal. We tested it on a fresh account with no card, and the run stopped with Modal's message, "Please add a payment method to use L4 GPU functions." No GPU started, nothing downloaded and nothing was billed. That is Modal's rule, not the agent's.
For Runpod, Vast.ai and Colab, their pricing pages did not state a card requirement, so we are not claiming either way. Treat an empty cell as unknown rather than as a "no".
Is there a free tier?
Modal has a monthly credit. It is a credit, not a free tier.
Modal's Starter plan is $0 per month plus compute, and it includes $30 of credits each month. Read that as an allowance. It applies because a card is on file, it resets, it can change, and it runs out. On our account the usage limit was one we had set ourselves, and we do not know what a brand-new account starts with.
The pricing pages of Runpod, Vast.ai and Colab did not state a comparable credit when we read them, so we are not listing one. If a platform advertises free GPU time, read the conditions on its own page; a free allowance that requires a card and resets monthly is a different thing from a free tier.
What is not included in the GPU price?
Everything that is not the graphics card, plus the time the machine is up but not generating.
- The processor and the memory are separate line items. Modal's pricing page lists CPU at $0.0000131 per core per second and memory at $0.00000222 per GiB per second, on top of the GPU rate. On our run those two together were roughly a third of the total.
- Billed time starts before the first image and ends after the last. The container has to come up and the model has to load before anything is generated, and you pay for that window. Short jobs pay proportionally more of it.
- Storage is a separate charge. Modal lists volumes at $0.09 per GiB per month. Keeping the weights stored is what makes the second run cheap, and it is not free either.
- Network egress is a separate charge. Modal lists it at $0.04 per GiB, with 1 TiB per month included on the Starter plan.
- The list price is not your bill. A price per hour or per second tells you the rate. The bill depends on how long the machine ran and what else was attached to it.
Why is a second image cheaper than the first?
Because the weights download once, and the download is most of the first run.
The Krea 2 Turbo weights we used come to about 18.6 GB across three files. On a cold machine, pulling those down is the bulk of the wall-clock time, and the machine is billing while it happens. Put them in attached storage (Modal calls it a Volume) and every later run skips the download entirely.
Our quoted $0.024 is a cached rerun, not a cold start. The first full session took about 13 minutes from the earliest recorded frame to the finished image, though most of that was a person approving things rather than compute.
How long does one image take?
About 95 seconds of execution on the run we measured, and about 3 minutes from sending the prompt to the image landing.
The gap between those two numbers is the machine starting and the model loading, not the sampling itself. The agent reported its own stage timings: ComfyUI ready in about 25 seconds, the image in about 50 seconds, the container up for about 75 seconds. Treat the generation time as the small part and the setup as the part that decides whether a job feels slow.
If you need a number for your own workload, run one image and read the settled billing, rather than multiplying a rate by a guess at the duration.
What should I check before I run anything?
Set the spending limit first. It is the one part of this you should not delegate.
- Open the billing page and find both controlsModal shows a usage limit and a separate, harder spend limit on the same page. They are two different settings. Read both before you run anything.
- Choose the numbers yourselfWe set our usage limit to $30, matching the monthly credit, and left the spend limit at the account's own default rather than inventing a value. We have not tested exactly what each control stops.
- Run one image, then read the settled billRun a single image and wait for billing to settle. Then open the platform's own billing page. An agent's summary of the cost is an estimate until the platform settles it.
Which platform should you pick?
We cannot rank them from a single measured run, and we are not going to guess. Here is what the same-day list prices actually say.
On the rates above, Vast.ai's low end is the cheapest per hour for an L4 at $0.25, but it is a marketplace: the $0.25 is the cheapest live offer, the median is $0.32, and the machine you get is someone else's. Modal's L4 works out to about $0.80 an hour from its per-second rate, and it bills per second with a serverless model, so an idle container is not left running against you in the same way. Runpod's L4 sits at $0.49 an hour on Community Cloud, with a serverless option at $0.69 an hour.
None of those headline figures includes the processor, the memory or the time the machine sits idle, and those are exactly the parts that made up roughly a third of our own bill. Pick on how the platform bills and how it handles idle time, then measure your own first image.
Not affiliated with or sponsored by Modal, Runpod, Vast.ai, Google or Krea. We use our own accounts and pay the platforms directly.
Where these numbers come from
Every figure on this page, with the page it was read from and the date it was read.
- Modal list price: $0.000222 per second for an Nvidia L4, read from modal.com/pricing on 2026-10-05. The page also lists CPU at $0.0000131 per core per second and memory at $0.00000222 per GiB per second.
- Modal card requirement: modal.com/docs/guide/billing, read on 2026-10-05. The page states that a payment method must be on file in order to use Modal.
- Runpod list price: $0.49 per hour for an L4 on Community Cloud, read from runpod.io/pricing on 2026-10-05. Their page carries the date 2026-09-27.
- Vast.ai list price: L4 from $0.25 per hour with a median of $0.32 per hour, read from vast.ai/pricing on 2026-10-05. Their page states that prices update hourly.
- Our measured usage: about $0.024, from Muse's settled report on 2026-10-03, in the same table published on the episode 2 page. L4 about $0.0173, CPU about $0.0056, memory about $0.0012.
- Conversion we did ourselves: $0.000222 per second times 3,600 seconds gives about $0.80 per hour. That is our arithmetic, not a figure from Modal's page. It lines up with the L4 rate quoted on the episode 2 page, which was read on 2026-09-30.
- Not verified here: whether Runpod, Vast.ai and Colab require a card, and whether they offer free credit. Their pricing pages did not say, so those cells read "Not verified" rather than a guess.
Questions we get asked
Short answers, same sources as above.
How much does one ComfyUI image cost on a cloud GPU?
We measured about $0.024 in usage for one image on a Modal L4, reported by the platform's own billing after it settled. That was a cached rerun, so the model weights were already stored and only the run itself was billed. Your number will differ with the model, the resolution and the step count.
Is the GPU the whole cost?
No. The graphics card is usually the smaller share than people expect. On our run the L4 portion was about $0.0173 of the $0.024, and the rest went to CPU and memory. Storage and network egress are billed separately again.
Do I need a credit card to use Modal?
Yes. Modal's billing documentation says you must have a payment method on file to use Modal, and we confirmed it on a fresh account: the run stopped with the message "Please add a payment method to use L4 GPU functions." No GPU started and nothing was billed.
Is the $30 Modal credit a free tier?
No. It is a monthly credit on the Starter plan. It applies because a card is on file, it resets each month, it can change, and it runs out. Read your own billing page before you run anything.
Why is the first image more expensive than the ones after it?
Because the weights download once. The Krea 2 Turbo weights are about 18.6 GB. On a cold machine that download dominates the run. Store them in attached storage and later runs skip it.
Which of these platforms is cheapest?
We cannot answer that from one measured run, and we are not going to guess. We have measured exactly one platform. The list prices above are all read on the same day from each vendor's own page, and on those numbers Vast.ai's low end is the cheapest per hour and Modal's per-second L4 works out to about $0.80 an hour. What you actually pay depends on the time the machine sits idle, which none of these headline figures include.
What is not in the listed GPU price?
Idle time, the processor, the memory, attached storage and network egress. A container is up before the first image and stays up after the last one, and you pay for that too. Short jobs pay proportionally more of it.
See the runs behind these numbers
Episode 1 sets up ComfyUI and Krea 2 on a rented GPU from a laptop with no graphics card, and it comes with the free kit: the prompt, the skill file and the script. Read episode 1.
Episode 2 hands the same job to Meta's Muse, which works on its own cloud computer, and shows the settled bill at the end. Read episode 2.
Both episodes come with the same free kit. Get the kit.