RunPod
Listed at https://www.runpod.io
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Per-second GPU: H100 SXM $2.99/hr, A100 SXM $1.49/hr, B200 $5.89/hr, L40S $0.99/hr, RTX 4090 $0.69/hr | per million tokens |
| Output | Same as input (per-GPU-second — no idle fees, no egress fees) | per million tokens |
| Cache Read | N/A (billed on actual GPU runtime, no prompt cache concept) | Discounted |
🤖 Supported Models (22)
✨ Pros
- ✓Per-second GPU billing, zero idle fees, no egress — cost-leader for bursty/inference workloads
- ✓13+ GPU tiers from RTX 4090 ($0.69/hr) to B300 288GB ($7.39/hr) — widest catalog among serverless GPU clouds
- ✓Serverless FlashBoot sub-200ms cold start; Pods deploy in under 30 seconds across 31 global regions
- ✓Community Cloud (low cost, peer-supplied GPUs) + Secure Cloud (Tier-3 datacenter) tier choice
- ✓Hub ships 100+ ready-to-deploy OSS templates (Llama, Qwen, DeepSeek, ComfyUI, Whisper)
- ✓Referral program: lifetime 5% credit rebate on referred developers' spending
⚠️ Cons
- ×Not a model API provider — you bring your own model + container image
- ×Community Cloud GPUs are interruptible (~15% preemption rate/hour) — production should use Secure Cloud
- ×Volume reservations require 1-3 month commit; pay-as-you-go is 20-40% more expensive than reserved
- ×No native prompt cache — long-context repeated-prefix workloads cost more than Anthropic/OpenAI
- ×China access unstable — needs proxy; no China-direct edge
- ×Documentation is sparser than AWS/GCP; expect to read template JSON for debugging
🎯 Best For
Open-source LLM hosting with low idle-cost (ComfyUI / Llama / Qwen / DeepSeek); bursty inference workloads; developers who need sub-30s GPU deployment without Kubernetes overhead; budget-conscious AI startups
💰 Pricing & Plans
| Tier / GPU | Billing Unit | Rate | Best For |
|---|---|---|---|
| Signup credit | One-time balance | $0 (no advertised starter credit; prepaid top-up required) | Evaluation before committing spend |
| Community Cloud — RTX 4090 24GB | Per second GPU-time | $0.74/hr (~$0.000206/s) | Cheap fine-tuning, batch jobs, side projects |
| Community Cloud — A100 SXM 80GB | Per second GPU-time | $1.59/hr (~$0.000442/s) | 70B-class serving, mid-size training runs |
| Community Cloud — H100 SXM 80GB | Per second GPU-time | $3.49/hr (~$0.000969/s) | Frontier LLM inference and training |
| Community Cloud — B200 180GB | Per second GPU-time | $6.79/hr (~$0.001886/s) | Largest open models, multi-node pretraining |
| Community Cloud — B300 288GB HBM3e | Per second GPU-time | $7.89/hr (~$0.002192/s) | Top-end open-weight model research |
| Serverless Flex — H100 80GB | Per second, active worker only | $4.79/hr while worker runs | Bursty API inference behind an endpoint |
| Serverless Flex — L4 / A5000 / 3090 24GB | Per second, active worker only | $0.69/hr while worker runs | Small-to-medium model APIs, cost-optimised |
| Storage — container disk | Per GB / month (5-min intervals) | ~$0.10/GB/month | Ephemeral worker scratch space |
| Storage — network volume | Per GB / month | $0.07/GB/month (<1TB), $0.05/GB/month (>1TB) | Persistent shared data across workers |
| Serverless Active workers | Per hour, always-on | Custom — discounts via sales enquiry | Consistent low-latency production traffic |
| Clusters (multi-node, SLA) | Reserved capacity | Custom quote (discounts at 10,000+ GPUs) | Distributed training at enterprise scale |
🔧 API & Developer Experience
- •REST API v2: Base https://api.runpod.io/v2 with a published OpenAPI schema; modernised replacement for the legacy GraphQL and REST v1 surfaces
- •Official SDKs: Python runpod-serverless SDK for endpoint handlers plus runpod-flash for writing and deploying your own Python functions
- •runpodctl CLI: One binary for Pods, endpoints, templates, volumes, Hub deployments, file transfer, SSH keys and model caching (Homebrew-installable)
- •MCP servers (2): Hosted API MCP at https://mcp.getrunpod.io/ (Streamable HTTP, OAuth or API key) for infrastructure control, plus a no-auth docs MCP at https://docs.runpod.io/mcp
- •AI coding-agent plugin: Router-based skills bundle for Claude Code / Codex / Cursor: runpod-mcp, runpodctl, flash, companion-clis, runpod-usage, runpod-migrate
- •Billing & account safety: Prepaid credit balance deducted in real time every 5 minutes; default spend limit of $80/hour across all resources; low-balance alerts and auto-pay
- •Hub & templates: Runpod Hub ships preconfigured open-source apps (vLLM, ComfyUI, Llama, Qwen, Whisper) deployable in one click, with community and custom templates
⚡ On-Demand GPU Marketplace & Serverless FlashBoot
Runpod's defining trait is that it sells raw GPU-seconds rather than model tokens, so you rent capacity instead of an API. The catalog spans roughly twenty tiers from RTX A5000 at $0.27/hr and RTX 4090 at $0.74/hr up through A100, H100, H200, B200 and B300 288GB HBM3e at $7.89/hr, with each listing documenting VRAM, RAM and vCPU counts so sizing is explicit rather than hidden. Two capacity pools let you trade reliability for price: Community Cloud draws on peer-hosted hardware across thirty-plus regions, while Secure Cloud places the same GPUs in Tier-3 datacentres. Serverless endpoints add autoscaling workers, so idle capacity costs nothing and FlashBoot is claimed to cut cold starts to under 200 milliseconds. Multi-node Instant Clusters extend the same model to distributed training, and Runpod Hub supplies preconfigured templates for vLLM, ComfyUI, Llama, Qwen and Whisper.
🌐 China Access & Latency
Runpod runs no mainland China edge or ICP-licensed infrastructure and publishes no China-dedicated region, so mainland developers should assume they are routing internationally. Access to the console, the REST API v2 and Serverless endpoints generally works through the public internet, but throughput and latency fluctuate and a stable proxy is recommended for interactive work such as SSH sessions and large container pulls. Object traffic is the real cost driver: pulling multi-gigabyte model images from Docker Hub or Hugging Face over a cross-border link can dominate wall-clock time, so pre-cached templates and model caching pay off disproportionately. Raw GPU rates stay identical regardless of origin, and network transfer itself is not metered, so the penalty shows up as slower first-token latency and longer deploys rather than a bigger bill. Teams needing predictable latency typically keep weights on a network volume and favour Singapore or Tokyo capacity.