RunPod pricing in 2026: GPU Pods, Serverless, and storage

RunPod pricing depends on the product you use. GPU Pods provide dedicated instances, Serverless charges while workers start and run, Clusters cover coordinated multi-node jobs, and storage is billed separately. Compute and most storage charges use fine-grained billing, but the hourly figure on a pricing page is only the starting point for a budget.

As of August 20, 2026, RunPod’s public prices ranged from $0.27 per hour for an RTX A5000 Pod to $7.89 per hour for a B300 Pod. Serverless hourly equivalents ranged from $0.58 for a 16 GB worker class to $9.98 for B300. Rates can change with GPU supply, deployment type, region, and configuration, so confirm the final figure on the live RunPod pricing page or in the deployment console.

RunPod pricing: the short version

RunPod product	How billing works	Best fit	Costs to check
GPU Pods	GPU compute billed at the displayed rate with fine-grained usage billing	Development, training, notebooks, batch jobs, and long-running services	GPU, container disk, volume disk, network volume, and stopped storage
Serverless	Per second from worker start until full stop, rounded up to the nearest second	Request-driven inference and traffic that changes over time	Startup, execution, idle timeout, worker storage, and network storage
Clusters	Per-second on-demand pricing; reserved capacity is quoted separately	Distributed training, large batch workloads, and jobs spanning multiple nodes	GPU count, node count, runtime, shared storage, and reservation terms
Storage	Per-second or hourly metering depending on storage type	Containers, checkpoints, shared datasets, and persistent workload data	Running versus stopped rates, capacity, persistence, and backups

Current RunPod GPU Pod prices

Pods are dedicated GPU environments that give you control over the container, storage, GPU type, and runtime. The table below is a dated sample from RunPod’s public pricing page on August 20, 2026. It is useful for planning, but it is not a permanent rate card.

GPU Pod	VRAM	Public hourly rate checked
RTX A5000	24 GB	$0.27/hour
A40	48 GB	$0.44/hour
RTX 4090	24 GB	$0.74/hour
A100 PCIe	80 GB	$1.39/hour
H100 PCIe	80 GB	$2.89/hour
H200	141 GB	$4.59/hour
B300	288 GB	$7.89/hour

The GPU name is not enough for a fair provider comparison. Check whether the price covers one GPU or several, the PCIe or SXM variant, CPU and RAM allocation, storage, network performance, region, interruptibility, and availability. A single-GPU listing should not be compared directly with an eight-GPU hyperscaler instance.

On-demand Pods and savings plans

RunPod’s Pod pricing documentation describes two main purchasing options:

• On-demand: pay as you go at the standard displayed rate, with no term commitment. RunPod requires enough prepaid credit to cover at least one hour of the selected configuration before deployment.
• Savings plan: prepay a three- or six-month term for a discounted compute rate on a specific GPU type. The plan is non-refundable, expires on a fixed date, and covers GPU compute only. Storage continues at standard rates.

A savings plan makes sense when the workload uses the same GPU type consistently enough to justify prepayment. Calculate expected active hours instead of assuming that a discounted rate produces a lower total bill. If utilization is uncertain, on-demand capacity keeps the commitment risk lower.

RunPod also supports interruptible capacity in parts of its platform. Treat that as a workload decision, not just a price decision. Training and batch jobs need frequent checkpoints, restart logic, and storage that survives an interruption.

How RunPod Serverless pricing works

RunPod Serverless is designed for containerized inference behind an API. According to the Serverless pricing documentation, billing starts when a worker starts and ends when it fully stops, rounded up to the nearest second.

There are two worker behaviors:

• Flex workers can scale to zero when idle. They suit variable demand, but each new worker may incur startup and model-loading time.
• Active workers remain running for consistent traffic and low-latency requirements. Because they are always on, they can accrue charges even when request volume is low. Discounts may be available through sales.

The bill includes more than request execution. RunPod charges for worker startup, model loading, request processing, and the configured idle period before a worker scales down. Container disk and network volumes are separate cost components.

RunPod’s public page listed these Serverless hourly equivalents on August 20, 2026:

Serverless GPU class	Public hourly equivalent checked
16 GB class	$0.58/hour
24 GB class	$0.69/hour
RTX 4090, 24 GB	$1.10/hour
A100, 80 GB	$2.72/hour
H100, 80 GB	$4.79/hour
B300, 280 GB worker class	$9.98/hour

For sporadic inference, Serverless can avoid paying for an idle Pod. For steady traffic, compare an active worker, a dedicated Pod, and your own autoscaling setup using measured request duration and utilization. The cheapest option depends on the traffic pattern and model-loading behavior.

RunPod Cluster pricing

RunPod Clusters are multi-node environments for distributed training, large batch jobs, and workloads that need coordinated GPU capacity. The Clusters product page describes on-demand, per-second billing with no minimum runtime. Reserved Clusters add guaranteed capacity, service commitments, and longer terms through a sales quote.

The public pricing page listed A100 SXM Cluster capacity at $1.79 per GPU-hour and H200 SXM at $4.31 per GPU-hour on August 20, 2026. Other cluster configurations may require a quote. Multiply the per-GPU rate by the GPU count and runtime, then include shared storage and any persistent resources.

RunPod storage pricing

Storage is where an apparently cheap GPU session can keep generating charges after compute stops. RunPod separates temporary container storage, Pod volume storage, and portable network volumes.

Storage type	Published rate	What to know
Container disk	$0.10/GB/month while running	Temporary and erased when the Pod stops
Volume disk	$0.10/GB/month running; $0.20/GB/month stopped	Persists through stops, but is deleted when the Pod is terminated
Standard network volume	$0.07/GB/month under 1 TB; $0.05/GB/month over 1 TB	Exists independently and can be moved or shared between compatible resources
High-performance network storage	$0.14/GB/month	Availability and location can be limited; confirm in the console

Container and volume disks use per-second storage metering; network volumes are billed hourly. RunPod’s documentation also warns that the platform is not intended as long-term archival storage. Back up important models, datasets, and checkpoints outside the account.

How to calculate your real RunPod cost

Use this formula before funding an account:

Total cost = compute runtime + worker startup and idle time + persistent storage + stopped storage + parallel resources.

For example, 100 hours on the public $0.74/hour RTX 4090 Pod rate is $74. A 200 GB standard network volume retained for a full month adds $14 at $0.07/GB/month, producing a simple $88 subtotal. That excludes other Pods, extra storage, tax, currency effects, and any change in the live rate.

For Serverless, estimate:

1. Average worker seconds per request, including model loading where applicable.
2. Requests per month and expected concurrency.
3. Idle timeout seconds that remain billable after work finishes.
4. Minimum active workers kept online continuously.
5. Container and network storage attached to the endpoint.

Test with representative traffic before committing to a worker configuration. A lower per-second price can still cost more if cold starts, long idle timeouts, or excessive worker counts increase billed time.

Billing details that are easy to miss

• RunPod uses prepaid credits. Charges are deducted from the account balance as resources run.
• Low balance can interrupt workloads. RunPod says Pods are stopped or terminated when the balance reaches zero, depending on whether a network volume is attached.
• Storage may continue after compute stops. Network volumes and stopped Pod volumes can keep accruing charges.
• Credits are non-refundable. Deposit only what you expect to use and configure balance alerts or auto-pay carefully.
• The default account spending limit is $80 per hour. Larger deployments may require a support request.
• RunPod documents no data-transfer fees. That removes an egress line item, but storage, runtime, and application-level transfer work still affect the total.

The current rules and low-balance behavior are documented in RunPod’s billing overview.

How to compare RunPod with another GPU cloud

Compare equivalent workloads and configurations, not isolated headline rates. Match the GPU model and count, VRAM, CPU, RAM, storage, network, region, availability model, interruption policy, billing granularity, and support level.

Our GPU pricing chart guide explains why spot, marketplace, fixed on-demand, and bundled hyperscaler prices should not be placed in one column without context.

For a narrower alternative, Hivenet currently publishes fixed on-demand rates of €0.40/hour for RTX 4090 and €0.75/hour for RTX 5090, with per-second billing. Those are different GPUs and a different currency from many RunPod listings, so compare the workload fit and complete configuration before comparing the numbers. See the current Hivenet GPU and CPU rental options and our RTX 4090 and RTX 5090 cloud pricing guide.

When RunPod pricing is a good fit

RunPod is worth evaluating when you need a broad GPU catalog, container control, request-driven Serverless inference, or self-service multi-node capacity. Pods suit interactive development and sustained jobs. Serverless suits variable inference traffic. Clusters suit workloads that must coordinate across nodes.

It is a weaker fit when you need long-term archival storage, cannot work with prepaid credits, require a specific GPU in a specific location without checking capacity, or cannot tolerate the operational risk of an interruptible instance.

Frequently asked questions

Does RunPod bill by the second?

RunPod’s billing documentation says compute and most storage charges use per-second metering, while network volumes are billed hourly. Serverless worker time is rounded up to the nearest second.

What is the difference between a RunPod Pod and Serverless?

A Pod is a dedicated GPU environment that you manage. Serverless runs containerized inference workers behind an endpoint and scales them according to demand. Pods fit development, training, and long-running services; Serverless fits request-driven inference.

Does RunPod charge egress fees?

RunPod’s current billing documentation says it does not charge for data transfer. Storage and the compute work involved in processing or moving data still have costs.

Do stopped RunPod Pods still cost money?

GPU compute stops, but persistent volume storage can continue. RunPod lists volume disk at $0.20/GB/month while a Pod is stopped. Network volumes also continue to accrue storage charges.

Are RunPod savings plans refundable?

No. RunPod describes three- and six-month savings plans as prepaid, non-refundable commitments with fixed expiration dates. They cover GPU compute for the selected GPU type, while storage is billed separately.

What is the cheapest RunPod GPU?

The answer changes with inventory and deployment options. On August 20, 2026, the public Pod pricing page listed RTX A5000 capacity at $0.27/hour among the displayed options. Check the live console because availability and rates can change.

Pricing checked against RunPod’s public pricing and documentation pages on August 20, 2026. Treat every figure as a dated snapshot and verify the live deployment total before purchasing credits or committing to a savings plan.