GPU cloud and LLM API prices, side by side
On-demand prices for H100, H200, B200, A100 and L40S, plus per-token API prices for popular open models, all checked against each provider's own pricing page. Updated 2026-10-12.
Cheapest GPU per hour
| GPU | Memory | Lowest $/hr | Cheapest at | Type | Highest $/hr | Providers |
|---|---|---|---|---|---|---|
| NVIDIA H100 | 80 GB | $2.20 | DeepInfra | Managed endpoint | $8.00 | 12 |
| NVIDIA H200 | 141 GB | $2.69 | DeepInfra | Managed endpoint | $8.00 | 8 |
| NVIDIA B200 | 180 GB | $3.69 | DeepInfra | Managed endpoint | $13.00 | 9 |
| NVIDIA A100 | 80 GB | $0.89 | DeepInfra | Managed endpoint | $4.00 | 7 |
| NVIDIA L40S | 48 GB | $1.09 | RunPod | Cloud VM | $2.25 | 5 |
"Cloud VM" means a raw instance you manage. "Serverless" is billed per second while code runs. "Managed endpoint" means the provider runs your model for you, so it costs more per hour. "Marketplace" prices are set by individual hosts and change by the minute.
Cheapest API price per model
| Model | Maker | Cheapest input | Cheapest output | Cheapest at (blended) | Providers |
|---|---|---|---|---|---|
| gpt-oss-120b | OpenAI | $0.10 | $0.50 | Baseten | 4 |
| gpt-oss-20b | OpenAI | $0.075 | $0.30 | Groq | 1 |
| GLM-5.3 | Z.ai | $1.40 | $4.40 | Tie: Together AI, Fireworks, Baseten | 3 |
| GLM-5.3 Flash | Z.ai | $0.15 | $0.50 | Tie: Together AI, Fireworks, Baseten | 3 |
| Kimi K3 | Moonshot AI | $2.85 | $14.25 | DeepInfra | 4 |
| DeepSeek V4 Flash (0731) | DeepSeek | $0.06 | $0.18 | DeepInfra | 2 |
| DeepSeek V4.1 Flash | DeepSeek | $0.30 | $1.20 | Tie: Fireworks, Baseten | 2 |
| Nemotron 3 Ultra | NVIDIA | $0.50 | $2.20 | DeepInfra | 3 |
| Qwen3 235B A22B Instruct 2507 | Alibaba | $0.09 | $0.55 | DeepInfra | 2 |
| Llama 3.3 70B | Meta | $0.10 | $0.32 | DeepInfra | 2 |
Prices are per 1M tokens. "Blended" weights input and output 3:1, which is typical for chat workloads.
Self-host or use an API? Break-even calculator
Work out what one million output tokens costs on a GPU you rent, then compare it with an API price. Use throughput from your own benchmark. It depends on the model, batch size, context length and serving engine.
Example values only. Self-hosting also costs engineering time, idle capacity and on-call. Utilization is the share of paid GPU hours spent serving real traffic.