LLM Cost Calculator: 3-Year Total Cost of Ownership for GPT, Claude, Gemini and Self-Hosting
Token prices are the number vendors quote, but they are rarely the biggest cost of running a language model in production: engineers, retrieval and monitoring often cost more. Enter your volume once and this calculator ranks ten current API models and a self-hosted option by total monthly cost, 3-year cost and cost per request, using each vendor's published list prices.
The calculator
Results update as you type, and the cheapest option moves to the top. Nothing you enter leaves your browser.
List prices per 1 million tokens, standard tier, short prompts, US dollars, from each vendor's official pricing page. Batch processing is 50% off at all three vendors. Prices last verified 11-08-2026.
How the calculation works
- Model cost = (uncached input tokens x input price + cached input tokens x cached price + output tokens x output price) / 1,000,000, less 50% on the share sent as batch jobs.
- Platform and people = engineers x loaded annual cost / 12, plus retrieval, plus observability, evaluation and hosting. It is the same for every hosted model.
- Self-hosted = GPUs x 730 hours x price per GPU-hour, plus the extra engineers to run them, plus the same platform and people cost.
- 3-year cost = monthly total x 36, with no discounting or price changes. Cost per request = monthly total / requests.
It compares cost only. A cheaper model that fails more often costs more in rework and escalations, so shortlist on price and decide on results from your own evaluation set.
Current list prices
Standard-tier prices for short prompts. Long prompts, priority processing and regional endpoints cost more; enterprise agreements often cost less.
| Model | Vendor | Input | Cached input | Output | Source |
|---|---|---|---|---|---|
| GPT-5.6 Sol | OpenAI | $4.00 | $0.40 | $20.00 | OpenAI pricing |
| GPT-5.6 Terra | OpenAI | $2.00 | $0.20 | $12.00 | OpenAI pricing |
| GPT-5.6 Luna | OpenAI | $0.20 | $0.020 | $1.20 | OpenAI pricing |
| GPT-5.4 mini | OpenAI | $0.75 | $0.075 | $4.50 | OpenAI pricing |
| Claude Fable 5.1 | Anthropic | $10.00 | $0.25 | $50.00 | Anthropic pricing |
| Claude Opus 5.5 | Anthropic | $4.00 | $0.20 | $20.00 | Anthropic pricing |
| Claude Sonnet 5.5 | Anthropic | $2.00 | $0.10 | $10.00 | Anthropic pricing |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.010 | $0.50 | Anthropic pricing |
| Gemini 3.1 Pro (preview) | $2.00 | $0.20 | $12.00 | Google pricing | |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | Google pricing |
US dollars per 1 million tokens, last verified 11-08-2026.
What changes the answer
- Volume. At low volume, people and platform dominate and the model barely matters. At millions of requests a month, token prices decide the ranking.
- Output length. Output tokens cost five to ten times as much as input tokens at most vendors, so long answers and agent reasoning add up fast.
- Caching and batch. Reused system prompts and documents bill at a fraction of the input price, and batch jobs are half price at all three vendors123.
- Tokenizers differ. The same text can produce noticeably different token counts on different vendors' models, so measure token use with your own prompts before comparing.
- Self-hosting. A GPU node costs the same whether it is busy or idle. It pays off only at high, steady volume, or when data must stay on infrastructure you control. Hourly rates for an H100 GPU range from about $4.59 at a specialist cloud to $7 to $12 on hyperscaler on-demand pricing67.
To weigh model cost against the business value of a workflow, use the ROI calculator; to compare who could build it, see the comparison matrix.
Where the defaults come from
| Default | Value | Basis |
|---|---|---|
| Model prices | Per model | Each vendor's official pricing page123 |
| Loaded engineer cost | $190,000 a year | Median software developer pay of $135,980 (BLS, May 2025)4 times 1.4 for benefits and overhead, an assumption |
| Retrieval | $500 a month | Pinecone's Enterprise plan minimum; its Standard plan starts at $505 |
| GPU price | $4.59 per GPU-hour | Lambda's listed rate for an 8x H100 node6 |
| Usage, engineers, other platform costs | As shown | Assumptions for a mid-sized internal assistant; replace with your own |
Frequently asked questions
What is LLM total cost of ownership?
Everything it costs to run a language model in production: tokens or GPUs, plus the engineers who build and operate it, retrieval, monitoring and evaluation. For many workloads the people cost more than the tokens.
Is self-hosting an open-weight model cheaper than an API?
Only at high, steady volume. A GPU node costs the same around the clock, and running it needs extra engineers. Below that volume, API pricing is usually cheaper.
How current are the prices?
Every page is validated roughly twice a week or more, and the price table shows its last-verified date. Check the vendor's page before you sign.
Does the calculator include enterprise discounts?
No. It uses public list prices. Enter a negotiated price by adjusting usage or treat the result as a ceiling.
References
- OpenAI, API pricing, accessed 11-08-2026. developers.openai.com
- Anthropic, Claude pricing, accessed 11-08-2026. platform.claude.com
- Google, Gemini API pricing, accessed 11-08-2026. aistudio.google.com
- U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, Software Developers. bls.gov
- Pinecone, pricing, accessed 11-08-2026. pinecone.io
- Lambda, GPU cloud pricing, accessed 11-08-2026. lambda.com
- Neurohive, "Cloud GPU pricing" guide (hyperscaler H100 on-demand rates). neurohive.io