How much does Runpod Serverless cost compared with a pod? The per-hour gap and the break-even

Runpod Serverless costs more per hour than a pod, but it only bills the seconds a worker is running. On Runpod's pricing page (read November 3, 2026), a serverless H100 flex worker is $5.39 an hour; an H100 pod is $1.99 an hour (H100 PCIe, Community Cloud) to $3.99 an hour (H100 SXM, Secure Cloud). A pod bills every hour it's on, busy or not. Serverless scales to zero, so it's the cheaper choice whenever your GPU is genuinely busy for fewer than about 9 to 18 hours a day, depending on which pod you'd otherwise rent.

How much does Runpod Serverless cost per hour?

From $0.58 an hour for a 16GB worker to $11.99 for a B300. These are Runpod's published flex rates (workers that scale to zero), read on its pricing and Serverless pages on November 3, 2026. You pay the same rate per second: the hourly figure divided by 3,600, so an H100 worker costs about $0.0015 for every second it runs.

The full flex rate card, with the closest pod on the same pricing page next to it:

Is Runpod Serverless cheaper than a pod?

Only if your GPU sits idle a good part of the day. A pod bills every hour it's switched on. Serverless bills per active second at a higher rate. So the deciding number is how many hours a day your endpoint is really busy, counting cold starts and idle-timeout seconds.

The break-even is simple: busy hours per day = 24 × pod rate ÷ serverless rate. Below that number, serverless is cheaper. Above it, the pod is.

Our arithmetic from the rates in the table above (Runpod's pricing page, November 3, 2026). It assumes you'd otherwise leave a pod running all day. If you'd switch a pod off overnight, the break-even moves down in proportion.

What does that look like in a month?

Two worked examples at the November 3 rates, 30 days:

• An H100 endpoint busy 2 hours a day (60 hours, including cold starts and idle seconds): 60 × $5.39 = $323.40 on Serverless. The same H100 as a pod left on: 720 hours × $1.99 = $1,432.80 (Community PCIe) or 720 × $3.99 = $2,872.80 (Secure SXM).
• A small image or speech model on a 24GB worker, 30 minutes a day (15 hours): 15 × $0.75 = $11.25 on Serverless, against 720 × $0.22 = $158.40 for an RTX 3090 Community pod left on.

Most side projects, demos and internal tools are nowhere near 9 busy hours a day. A production API that's busy from morning to night usually is, and a pod (or several) wins there. Not sure which GPU your model needs? Check it in our VRAM calculator before you pick a tier.

What does Runpod Serverless bill you for?

Compute while a worker runs, plus storage. Runpod's Serverless pricing docs say you're "billed from when a worker starts until it fully stops, rounded up to the nearest second", and list three billed phases.

Storage on top of compute, per Runpod's docs and pricing page (read November 3, 2026):

• Container disk: about $0.10/GB a month, charged in 5-minute intervals.
• Network volume (shared storage for model files): $0.07/GB a month under 1TB, $0.05/GB over 1TB; the high-performance tier is $0.14/GB a month.

What makes a Runpod Serverless bill bigger than you expect?

Five things, all from Runpod's own docs:

• A long idle timeout. The docs say a worker "is billed while it waits" for the next request. The console default is 5 seconds (10 seconds if you create the endpoint through REST v2). Raise it a lot to dodge cold starts and a scale-to-zero endpoint starts billing like an always-on one.
• Active workers. Setting active workers to 1 or more removes cold starts, but the docs say "Active workers incur charges continuously, including when they have no jobs." Runpod's pricing docs list active-worker discounts as "available through sales inquiry", with no public rate, so price them at the flex rate unless sales quotes you lower.
• Slow cold starts. You pay from the moment a worker starts until your handler is ready, which is mostly model loading. A big model that takes a minute to load bills that minute on every scale-up. Runpod says FlashBoot (on by default in the console, off by default through REST v2) and cached models shorten this. Runpod's Serverless page claims "sub-200ms FlashBoot cold starts"; that's Runpod's figure, we haven't measured it.
• Runaway jobs. The console's default execution timeout is 600 seconds per job. Keep a timeout on so a stuck job can't bill for hours.
• Rounding. Each worker's time is rounded up to the next second, so many tiny requests carry a little overhead each.

Two safety nets work in your favour: Runpod sets a default spend limit of $80 an hour across all resources, and an endpoint with no requests for 7 days has its max workers set to 0 (it's cut to 2 after 3 days, with an email). That second one can also surprise you: raise max workers again in the console before you use an old endpoint.

When should you rent a pod instead?

When the GPU is busy most of the day, or when the work isn't request-shaped:

• Steady traffic above the break-even in the table above (for an H100, about 9 busy hours a day against a Community PCIe pod).
• Training, fine-tuning and batch jobs that run for hours anyway. You'd pay the higher serverless rate for nothing.
• Interactive work (a notebook, ComfyUI, testing models by hand), where a pod you start and stop yourself is simpler.

On the pod side, remember a stopped pod isn't free either: Runpod's pricing page lists volume disk at $0.10/GB a month while running and $0.20/GB a month while idle. For the pod decision, see the cheapest H100 rental, Runpod vs Vast vs Lambda and our Runpod review.

The short version: Serverless is a bet that your GPU is idle most of the day. Count your real busy hours, check them against the break-even table, and see today's pod prices on our cloud-GPU compare page. For the bigger picture across providers, read the cloud-GPU pricing guide.