Free cost planning tool
GPU vs API cost calculator: local LLMs or a cloud API?
Estimate what your team’s token volume costs on a hosted API, then compare it with buying GPUs and serving open models through gpuOS: amortized cards, electricity and the plan fee. Prices checked 2026-10-27. No account required.
Your workload and hardware
USD per 1M input / output tokens, list prices checked 2026-10-27.
Board power from NVIDIA’s specifications.
Whole team, every day of a 30-day month. 20 people × 50 requests = 1,000.
Prompt, system prompt, chat history and retrieved context.
Generated tokens, including any reasoning tokens.
Cards you buy and connect to gpuOS.
Prefilled with a dated market price. Enter your quote.
Default: U.S. commercial average, rounded. Convert €/kWh first.
Cards are assumed to draw full board power for these hours.
Period over which the purchase is spread.
Community includes the base catalog only.
Estimated monthly difference
$315
saved per month compared with OpenAI GPT-6.1 Terra.
- Cloud API (GPT-6.1 Terra)
- $525
- Local total, per month
- $210
- GPUs amortized ($5,581 over 36 months)
- $155
- Electricity (173 kWh)
- $25.88
- gpuOS plan
- $29.00
- Break-even on the purchase
- 11.9 months
Average output needed: 55.6 tokens/s across GeForce RTX 5090 32 GB × 1, over 10 h a day.
Peaks are higher than the average. Measure your model’s throughput on this GPU before relying on the result; one card serves a limited number of concurrent requests.
USD, excluding VAT. Cards only: the host machine, cooling, idle power and your time are not included.
How the estimate works
Monthly tokens equal requests per day × 30 × tokens per request. The cloud cost multiplies input and output tokens by the provider’s standard list price per million tokens. Cached-input discounts, batch rates and long-prompt surcharges are not applied.
The local cost adds three parts: the purchase price of the cards divided by the amortization period; electricity, counted as each card’s maximum board power for the hours you enter, every day; and the gpuOS plan. Pro costs $29 per GPU per month, or $290 per GPU per year. Community is free for one GPU and the base catalog.
Break-even is the purchase price divided by the monthly API bill minus electricity and the plan. When those running costs already exceed the API bill, the cards never pay for themselves at that volume, and the calculator says so.
Cloud API reference prices
Standard list prices in USD per million tokens, read on each provider’s pricing page on 2026-10-27. Providers change prices often; check the source before you decide.
| Model | Input | Output | Source and notes |
|---|---|---|---|
| OpenAI GPT-6.1 Terra | $1 | $5 | developers.openai.com, checked 2026-10-27. Prompts up to 272K tokens; longer prompts cost $2 / $7.50. |
| OpenAI GPT-6 Luna | $0.1 | $0.5 | developers.openai.com, checked 2026-10-27. Prompts up to 272K tokens; longer prompts cost $0.20 / $0.75. |
| Anthropic Claude Sonnet 5.5 | $2 | $10 | claude.com, checked 2026-10-27. |
| Anthropic Claude Haiku 5.5 | $0.1 | $0.5 | claude.com, checked 2026-10-27. Prompts up to 100K tokens; longer prompts cost $0.50 / $2.50. |
| Google Gemini 3.8 Flash | $0.75 | $3.75 | ai.google.dev, checked 2026-10-27. Price through December 31, 2026; Google lists $1.50 / $7.50 from January 1, 2027. |
| Fireworks Open-weights dense model above 16B (32B–70B class) | $0.9 | $0.9 | docs.fireworks.ai, checked 2026-10-27. Fireworks' size-based rate for base models not priced individually. |
| Fireworks Open-weights dense model, 4B–16B | $0.2 | $0.2 | docs.fireworks.ai, checked 2026-10-27. Fireworks' size-based rate for base models not priced individually. |
| Fireworks Open-weights dense model under 4B | $0.1 | $0.1 | docs.fireworks.ai, checked 2026-10-27. Fireworks' size-based rate for base models not priced individually. |
| Fireworks Open-weights MoE model up to 56B | $0.5 | $0.5 | docs.fireworks.ai, checked 2026-10-27. Fireworks' size-based rate for base models not priced individually. |
| Fireworks Open-weights MoE model, 56.1B–176B | $1.2 | $1.2 | docs.fireworks.ai, checked 2026-10-27. Fireworks' size-based rate for base models not priced individually. |
| Fireworks gpt-oss-120b | $0.15 | $0.6 | docs.fireworks.ai, checked 2026-10-27. Together and Groq list the same $0.15 / $0.60. |
| Groq gpt-oss-20b | $0.075 | $0.3 | console.groq.com, checked 2026-10-27. |
Llama 3.3 70B is no longer listed as a serverless model on Together or Fireworks, so the 70B-class reference is Fireworks’ published rate for dense open-weights models above 16B parameters. In the gpuOS dashboard, usage is priced at the same model on a hosted provider when one lists it (gpt-oss), otherwise at the Fireworks rate for the same size and architecture.
GPU assumptions
GPU prices rose far above launch prices in 2026, so the calculator starts from dated market observations. Replace them with your own quote. Electricity defaults to $0.15 per kWh, rounded from the EIA U.S. commercial average of $0.1447 for August 2026 (preliminary).
| GPU | Max board power | Indicative price | Source |
|---|---|---|---|
| GeForce RTX 4090 24 GB | 450 W | $2,977 | gpupoet lowest average price, September 2026 (eBay listings; launch MSRP $1,599), checked 2026-10-27. |
| GeForce RTX 5090 32 GB | 575 W | $5,581 | gpupoet lowest average price, September 2026 (eBay listings; launch MSRP $1,999), checked 2026-10-27. |
| RTX PRO 6000 Blackwell 96 GB | 600 W | $16,251 | gpupoet lowest average price, September 2026 (NVIDIA Marketplace listed $16,000 in August 2026), checked 2026-10-27. |
| H100 NVL 94 GB (PCIe) | 400 W | $39,774 | Single-card reseller listing (Cloudninjas); OEM channel prices are higher, checked 2026-10-27. |
What the number leaves out
- Quality. A 32B open model on one card is not equivalent to a frontier API. Compare against the tier whose answers you would actually accept, and test your own prompts before switching. The hosted open-weights rows are the closest like-for-like comparison.
- Throughput. One GPU serves a limited number of concurrent requests, and Ollama processes a bounded number in parallel. The calculator shows the average output rate your volume needs; peaks need more. Add cards or spread the hours if your measured throughput falls short.
- Model fit. The model and its context must fit in VRAM. Check with the VRAM calculator.
- Everything around the cards. The host machine, power supply, cooling, rack space or hosting, idle power, spare parts and the time to run them are not included. Resale value at the end of the period is not deducted either.
- Data path. Inference runs on your GPUs, but request and response payloads pass through the hosted gpuOS gateway.
For a broader comparison of privacy, latency and operations, read local AI vs cloud AI or see gpuOS pricing.