Rent a GPU or pay per token? Break-even report
We earn commissions when you shop through the links below.
Pick a model and enter how many tokens you use a month. We check whether it fits a rentable GPU, work out the true monthly rental cost, compare it with the cheapest public API price we verified for the same model, and show the monthly token count where renting starts to cost less.
Rent a GPU or pay per token?
Estimates, not guarantees. Prices move hourly. Verify at source.
Excludes Turbo / promo / off-peak / Spot prices. Standard price only, exact model listing only.
Rent through: Runpod Referral link · Vast.ai Referral link · DigitalOcean Affiliate partner · Novita AI (API) Referral link
GPU prices checked 2026-11-09; API prices checked 2026-11-09.
Weights + KV cache from real Hugging Face file sizes, 90% fit rule, every number sourced.
How we calculate + sources ⓘ
How we calculated
- Fit. The same engine as our VRAM calculator: weights plus KV cache under 90% of GPU memory is Fit, up to 100% at 4-bit is Tight, more is Won't fit. With "Cheapest that fits" we try 1, 2, 4 and 8 of every GPU in our price table (Vast.ai: single GPUs only) and keep the cheapest setup.
- True monthly rental. On-demand $/GPU/hr from the GPU prices table × hours per month (default 730), plus storage where the provider publishes a rate, using the same rules as True monthly cost.
- API price. The lowest standard per-token price we could verify today on each provider's own pricing page for the exact same model. No promotions, off-peak, spot, batch or cached-input rates. Renamed provider variants (for example "-Turbo") are left out. Models without a verified price show No public API price.
- Blended price. APIs charge input and output tokens differently. We assume 3 input tokens for every 1 output token (75% / 25%) and show both prices so you can redo the sum for your own mix.
- Break-even = monthly rental ÷ API price per token. Below it the API is cheaper; above it renting is cheaper.
- What this leaves out. Whether one rented setup can actually serve your volume (throughput depends on the serving stack and batch size; see speed), your setup and ops time, and idle hours you still pay for. Treat the result as a first filter.
Sources ⓘ
Sources
Cheapest standard API price per model among the provider pricing pages we checked (Novita AI, DeepInfra, Together AI, Fireworks AI, DeepSeek). Each provider name links to the page we read; times are Pacific:
| Model | Cheapest API we found | $ in / out per 1M | Blended per 1M | Verified |
|---|---|---|---|---|
| DeepSeek-R1-0528-Qwen3-8B | Novita AIDeepSeek R1 0528 Qwen3 8B | $0.06 / $0.09 | $0.0675 | |
| DeepSeek-V3.2 | DeepInfraDeepSeek-V3.2 | $0.26 / $0.38 | $0.29 | |
| DeepSeek-V4.1-Flash-1017 | DeepInfraDeepSeek-V4.1-Flash-1017 (list price) | $0.20 / $0.60 | $0.30 | |
| Gemma 4 26B A4B | DeepInfragemma-4-26B-A4B-it | $0.07 / $0.34 | $0.1375 | |
| Gemma 4 31B | Novita AIGemma 4 31B | $0.14 / $0.40 | $0.205 | |
| GLM-5.3-Flash | DeepInfraGLM-5.3-Flash | $0.15 / $0.50 | $0.2375 | |
| gpt-oss-120b | Novita AIGPT OSS 120B | $0.05 / $0.25 | $0.10 | |
| Kimi K3 | DeepInfraKimi-K3 | $2.85 / $14.25 | $5.70 | |
| Llama 3.1 8B Instruct | Novita AILlama 3.1 8B Instruct | $0.02 / $0.05 | $0.0275 | |
| Llama 3.3 70B Instruct | Novita AILlama 3.3 70B Instruct | $0.135 / $0.40 | $0.2013 | |
| Phi-4 14B | DeepInfraphi-4 | $0.07 / $0.14 | $0.0875 | |
| Qwen3 14B | DeepInfraQwen3-14B | $0.12 / $0.24 | $0.15 | |
| Qwen3 32B | DeepInfraQwen3-32B | $0.08 / $0.28 | $0.13 | |
| Qwen3.5 9B | DeepInfraQwen3.5-9B | $0.10 / $0.15 | $0.1125 | |
| Qwen3.5-122B-A10B | Novita AIQwen3.5 122B A10B | $0.40 / $3.20 | $1.10 | |
| Qwen3.6 35B-A3B | DeepInfraQwen3.6-35B-A3B | $0.10 / $0.95 | $0.3125 | |
| Qwen3.8 27B | Novita AIQwen3.8 27B | $0.42 / $3.00 | $1.065 |
GPU rental prices come from Runpod, Lambda and Vast.ai as listed on GPU prices, each with its own verified time. Groq's pricing page redirected to its homepage when we checked, so it is not included. See our methodology and disclosure.