RunPod Pricing Guide: Calculate Your GPU Cloud Costs
Direct answer: RunPod pricing has two models. Pods charge by the hour for dedicated GPU instances you control directly. Serverless charges only for the seconds your container actively processes a job. There are no egress fees for HTTP responses, VPN bandwidth is billed separately, and persistent network volumes cost $0.07 per GB per month. This guide breaks down every line item and shows how to estimate your exact monthly bill.
Pod pricing: all GPU types (2025 rates)
| GPU | VRAM | Secure Cloud | Community Cloud | Best for |
|---|---|---|---|---|
| RTX A4000 | 16 GB | $0.20/hr | $0.17/hr | Light inference, beginners |
| RTX A4500 | 20 GB | $0.27/hr | $0.22/hr | Mid-size models |
| RTX 3090 | 24 GB | $0.44/hr | $0.32/hr | SDXL, LLM 7B inference |
| RTX 4090 | 24 GB | $0.74/hr | $0.59/hr | Fastest consumer GPU |
| RTX A6000 | 48 GB | $1.09/hr | $0.79/hr | Large models, multi-user |
| A100 40GB | 40 GB | $1.19/hr | $0.89/hr | Training, batch inference |
| A100 80GB | 80 GB | $1.99/hr | $1.49/hr | 70B LLMs, large-scale training |
| H100 80GB | 80 GB | Custom | N/A | Enterprise, massive models |
Serverless pricing: pay per job, not per hour
Serverless charges for the time your container is actively executing a job, measured in seconds. There is no idle cost. Workers spin up from a cold state, process the request, and shut down. You configure max workers to limit concurrency and cost.
| GPU | Per-second rate | Equivalent per-hour | Best for |
|---|---|---|---|
| RTX A4000 | $0.000056 | $0.20/hr | Light APIs, testing |
| RTX 3090 | $0.000122 | $0.44/hr | Image gen, LLM inference |
| RTX 4090 | $0.000206 | $0.74/hr | Fast inference APIs |
| A100 40GB | $0.000331 | $1.19/hr | Production LLM APIs |
| A100 80GB | $0.000553 | $1.99/hr | Massive model serving |
Storage pricing
| Type | Price | Notes |
|---|---|---|
| Network Volume | $0.07/GB/month | Persistent, survives pod termination |
| Container Disk | Included | Ephemeral, lost when pod stops |
| Template Storage | $0.07/GB/month | Custom saved templates |
Real cost scenarios
Scenario 1: hobbyist Stable Diffusion (10 images/day)
A user generates 10 images per day on an RTX 3090 serverless endpoint. Each image takes 15 seconds of GPU time. Daily GPU time: 150 seconds = 0.042 hours. Cost: 0.042 × $0.44 = $0.018/day or $0.55/month.
Scenario 2: AI app API (1,000 requests/day)
An app serves 1,000 LLM inference requests per day. Each request averages 3 seconds on an A100 40GB. Daily GPU time: 3,000 seconds = 0.83 hours. Cost: 0.83 × $1.19 = $0.99/day or $29.70/month.
A persistent A100 pod running 24/7 would cost $856/month. Serverless saves 96% for this bursty workload.
Scenario 3: training a LoRA (20 hours total)
Fine-tuning a Stable Diffusion LoRA on an RTX 3090 for 20 hours. Cost: 20 × $0.44 = $8.80. Add $2 for storage. Total: $10.80 for a custom model.
Scenario 4: production API (always-on)
A production API needs <5ms latency and cannot tolerate cold starts. An RTX 4090 pod runs 24/7. Monthly cost: 730 hours × $0.74 = $540.20/month. Add $10 for persistent storage. Total: ~$550/month.
Cost comparison: RunPod vs competitors
| Provider | RTX 3090 | A100 40GB | A100 80GB |
|---|---|---|---|
| RunPod | $0.44/hr | $1.19/hr | $1.99/hr |
| Vast.ai | $0.35–$0.60/hr | $0.90–$1.50/hr | $1.50–$2.50/hr |
| Lambda Labs | N/A | $1.10/hr | $1.99/hr |
| Paperspace | N/A | N/A | $2.46/hr |
| AWS g5.xlarge | N/A | N/A | N/A |
| AWS p4d.24xlarge | N/A | N/A | ~$32.77/hr (8× A100) |
How to minimize your RunPod bill
- Use Serverless for intermittent workloads: If your GPU sits idle > 50% of the time, serverless is cheaper.
- Stop pods when done: An idle pod burns money. Set a habit of stopping immediately after work.
- Use Community Cloud: 20–40% cheaper than Secure Cloud if you can tolerate variable availability.
- Right-size your GPU: Do not rent an A100 for a task that fits on a 3090. Check VRAM requirements first.
- Delete unused network volumes: At $0.07/GB/month, a 500 GB forgotten volume costs $35/month.
- Monitor with alerts: Set daily spend limits in the dashboard and enable email alerts.
- Batch jobs: Process multiple items in one pod session instead of starting/stopping repeatedly.
Free credits and trials
RunPod offers $5 in starter credits for new accounts. This covers roughly 11 hours on an RTX A4000 or 2.5 hours on a 3090. Enough to test workflows. Serverless billing starts after credits are consumed. No automatic subscription; you add funds manually.
Serverless vs Pod: cost decision tree
| Question | If yes → | If no → |
|---|---|---|
| Is traffic bursty or unpredictable? | Serverless | Pod |
| Do you need <1s cold start? | Pod (keep warm) | Serverless |
| Is GPU idle > 50% of the day? | Serverless | Pod |
| Do you need SSH / interactive Desktop? | Pod | Serverless |
| Are you building an API for users? | Serverless | Pod + load balancer |
Related guides
- Cheapest GPU Cloud Providers
- RunPod Serverless Tutorial
- RunPod vs Vast.ai
- RunPod vs Paperspace
- RunPod Review
FAQs
How does RunPod serverless billing work?
You pay for the seconds your container is actively processing a job. No idle charges. Workers scale from zero automatically.
Is there a monthly minimum?
No. RunPod is purely pay-as-you-go. Add funds when needed. Credits do not expire.
What happens if I forget to stop a pod?
It keeps billing by the hour. Set auto-shutdown or monitor via the dashboard to avoid surprises.
Are there egress fees?
HTTP responses from your endpoint are free. Large data transfers out of the platform may incur bandwidth charges.
Can I get a refund?
Refund policies vary. Contact support for billing issues. Unused credits are typically not refundable.
Start with $5 free credit
Create an account, claim your free credits, and test any GPU type for free.
Try RunPod