How does RunPod billing work?

RunPod billing is prepaid: you add credit, and every running pod, disk and serverless worker is deducted from it per second, with billing running every 5 minutes. You need at least one hour of credit for your chosen setup to deploy, and at $0 RunPod terminates pods that have no network volume. You need about $1 of credit to deploy an RTX 5090 on secure cloud with our template’s disks: one hour of GPU plus disks, at October 2026 prices. GPU prices are in our RunPod for AI images guide.

How does RunPod charge you?

You buy credit up front, and RunPod takes it from your balance as you use things. Compute and storage are both billed per second, and there are no fees for moving data in or out. Credits are deducted in real time: billing runs every 5 minutes, and charges come off continuously while anything runs or sits on a disk.

The rules that matter, from RunPod’s billing and pod pricing docs, checked 2 Oct 2026:

Add credit on the Billing page after you sign up at RunPod. RunPod says you can start with as little as $10.

How much credit do you need to deploy?

At least one hour of whatever you’re about to deploy: the GPU’s hourly price plus your disks. If your balance is lower, RunPod won’t deploy and asks you to add credit or pick a cheaper GPU. With our templates’ disks (30 GB container, 100 GB volume), disk storage adds only about 2 cents an hour.

Disk math: 130 GB × $0.10 per GB per month is $13 a month, which is under 2 cents an hour. RunPod’s deploy summary shows the same thing: a per-hour total of GPU, container disk and persistent storage (RunPod docs). That’s the minimum, not a budget. A real session runs longer than an hour, so keep a few hours of credit on the account while you work.

What does per-second billing mean in practice?

You pay for the seconds a pod runs, not whole hours. Ten minutes on an RTX 5090 at $0.99 an hour is about 17 cents. Two hours is two hours. The first-boot model download counts too: the GPU is yours, and billed, from the moment the pod runs, even while the template is still fetching its models.

Paid image APIs come out of the same balance. The Dataset Maker’s paid engines (Nano Banana Pro and Seedream) run on RunPod’s Public Endpoints, which you call with your RunPod API key, and RunPod counts Public Endpoint spend in your account alongside pods, Serverless and storage (billing API docs). RunPod lists Nano Banana Pro Edit at $0.14 a photo at 1k or 2k and $0.24 at 4k, and Seedream 4.0 at $0.027 (checked 2 Oct 2026). The Dataset Maker asks Nano Banana Pro for 2k, so its photos cost $0.14 each.

To price a whole build, from face to dataset to LoRA, use the RunPod cost calculator.

What happens when your balance hits $0?

RunPod stops every running pod. Pods with a network volume keep their data on that volume. Pods without one are terminated, and RunPod says their data can’t be recovered. Network volumes keep billing while stopped, and if the balance stays at $0 the volume itself can eventually be deleted too.

Our templates use a normal volume disk, not a network volume. So for most people reading this, $0 means the dataset or LoRA you were training is gone. Two settings on the Billing page prevent it:

1. Low balance alert. Under Notifications, turn on Low balance alert and set a threshold. RunPod emails you when the balance drops below it.
2. Auto-pay (optional). Add a card, set a threshold and an amount, and RunPod tops you up when the balance gets low. It tries at most once an hour.

Better still, download your files at the end of every session so a $0 balance can’t cost you work.

What is the $80 an hour spend limit?

A safety cap on how fast your account can spend: by default $80 an hour across all your pods, serverless endpoints and storage combined. It’s there to stop a misconfigured job from draining your balance. It isn’t a monthly limit. RunPod raises it automatically as your account builds history, or support can raise it sooner if you ask.

For one person training one character, you’ll never get near it. The RTX PRO 6000 the course trains on is $2.09 an hour, and even the H100 SXM, the priciest card in our GPU table, is $3.49 an hour.

Don’t confuse it with a budget. We cap the free Discord face generator at about $80 a month of RunPod spend ourselves, but that’s our own limit on our own bot, not a RunPod setting.

How is Serverless billed differently?

A serverless endpoint has no pod sitting there. You pay per second from the moment a worker starts until it fully stops, rounded up to the second. That covers three phases: starting the worker and loading the model, running the job, and the idle timeout after it. With nothing running, you pay nothing for compute.

From RunPod’s serverless pricing (checked 2 Oct 2026): flex workers scale to zero when idle, active workers run 24/7 (RunPod offers discounts on them through sales), and the idle timeout defaults to 5 seconds.

Our private bot uses max workers 1, idle timeout 5 seconds and FlashBoot, so it costs $0 idle; the 1 to 2 minute wake-up after a break is billed.

What can go wrong?

• You stop instead of terminate. The GPU bill ends, the disk bill doesn’t: a stopped volume disk bills at $0.20 per GB per month. See stop vs terminate.
• You can’t deploy. You don’t have one hour of credit for that setup. Add credit or pick a cheaper card.
• Your balance hits $0 mid-training. The pod is terminated and the LoRA is gone. Training takes 2 to 3 hours on an RTX PRO 6000 at $2.09 an hour, so keep credit for the whole run and turn on the low balance alert.
• You add $100 to try it out. Credits can’t be refunded or withdrawn. Start small.
• You redeploy a slow pod over and over. Every new pod downloads its models again, on billed GPU time. Read the logs first: why a RunPod pod gets stuck.
• A small top-up on a prepaid card fails. RunPod says prepaid cards should deposit at least $100 per transaction. Use a regular card for small amounts.