LLM API and GPU cost calculator
Estimate what an AI feature will cost before you build it. Compare LLM API prices across every major model for your own traffic, or price cloud GPU rental by the hour, from each provider’s published price list.
Prices last verified between and against each provider’s published price list.See every source
Estimate your AI costs
All amounts in US dollars, from each provider’s published price list.
The starting figures are examples. Replace them with your own.
GPT-6 Astra estimate
- Per month
- $7,604
- Average month (365 ÷ 12 days)
- Per year
- $91,250
- Per 1,000 requests
- $25.00
- Per request
- $0.025
| Price | Tokens per day | Cost per day | Share |
|---|---|---|---|
| Input | 10,000,000 | $100.00 | 40% |
| Output | 3,000,000 | $150.00 | 60% |
Sources: developers.openai.com (checked )
Compare every model for this workload
Same requests and token counts on each model, cheapest first. Select a column heading to sort, or a model name to see its breakdown above. List prices are the standard rates for the price tier your prompt length falls in.
| vs cheapest | |||||||
|---|---|---|---|---|---|---|---|
| Prompts up to 100,000 input tokens | Anthropic | $0.10 | $0.01 | $0.50 | $0.25 | $76.04 | Cheapest |
| Prompts up to 272K input tokens | OpenAI | $0.10 | $0.01 | $0.50 | $0.25 | $76.04 | Cheapest |
| Mistral AI | $0.15 | n/a | $0.60 | $0.33 | $100.38 | 1.3× | |
| Google (Gemini API) | $0.30 | $0.03 | $2.50 | $1.05 | $319.38 | 4.2× | |
| Mistral AI | $0.70 | $0.07 | $2.10 | $1.33 | $404.54 | 5.3× | |
| Google (Gemini API) | $0.75 | $0.075 | $3.75 | $1.88 | $570.31 | 7.5× | |
| Prompts up to 272K input tokens | OpenAI | $1.00 | $0.05 | $5.00 | $2.50 | $760.42 | 10.0× |
| Anthropic | $2.00 | $0.10 | $10.00 | $5.00 | $1,521 | 20.0× | |
| Google (Gemini API) | $2.00 | $0.10 | $10.00 | $5.00 | $1,521 | 20.0× | |
| Prompts up to 272K input tokens | OpenAI | $2.00 | $0.10 | $10.00 | $5.00 | $1,521 | 20.0× |
| Anthropic | $4.00 | $0.20 | $20.00 | $10.00 | $3,042 | 40.0× | |
| Anthropic | $8.00 | $0.20 | $40.00 | $20.00 | $6,083 | 80.0× | |
| Prompts up to 272K input tokens | OpenAI | $10.00 | $1.00 | $50.00 | $25.00 | $7,604 | 100.0× |
NVIDIA H100 80GB (SXM) estimate
- Per month
- $40,179
- 730 billed hours
- Per year
- $482,150
- Per hour
- $55.04
- 8 GPUs billed
- Per 1M tokens
- n/a
- Needs a throughput figure
- Priced on AWS EC2 p5.48xlarge (US East (N. Virginia), Linux), 1 × 8 GPUs.
Sources: aws.amazon.com (checked ); b0.p.awsstatic.com (checked )
Every provider for the NVIDIA H100 80GB (SXM)
Monthly cost for 8 GPUs at 24 hours a day, cheapest on demand price first. Select a provider to see its breakdown.
| Provider | On demand | Spot | 1-year reserved |
|---|---|---|---|
| $15,710 | No self-serve price published for this option. | No self-serve price published for this option. | |
| $23,302 | No self-serve price published for this option. | No self-serve price published for this option. | |
| $26,806 | No self-serve price published for this option. | No self-serve price published for this option. | |
| $40,179 | Price not verified yet. | Price not verified yet. | |
| $71,774 | $13,264 | $45,935 | |
| Price not verified yet. | Price not verified yet. | Price not verified yet. |
Guides for developers and CTOs
- How to estimate LLM API costs before launch
A step-by-step method for pricing an LLM feature before it ships: define the request, measure tokens, model traffic, check price tiers and turn one guess into a low, expected and high range.
- Prompt caching and batch APIs: what they actually save
Why a large cache discount on the price list turns into a smaller saving on the bill, when batch processing fits, and how to work out the real saving for your own traffic.
- API vs self-hosted GPUs: the break-even math
When renting GPUs and running your own model costs less than paying per token: the break-even formula, the capacity ceiling, why idle time matters most, and when a reservation pays off.
Frequently asked questions
- How does the LLM API calculator work out a cost?
- It splits each request into price buckets (standard input, cached input, output, and the batch versions of each) according to your token counts and the shares you set for caching and batch. Each bucket is multiplied by the model’s published price per million tokens. The cost per request is then multiplied by your requests per day, and a month is taken as 365 ÷ 12 days.
- Where do the prices come from?
- Each figure comes from the provider’s own published price list: OpenAI, Anthropic, Google (Gemini API) and Mistral AI for API models, and AWS EC2, Microsoft Azure, Google Cloud, Lambda and Runpod for GPU rental. Every price is stored with its source link and the date it was last checked, shown on the methodology page.
- Why is a model or GPU option shown as not available?
- Because the provider does not publish that price, or we have not yet been able to verify it. We leave the option out rather than guess, so every figure you see can be traced to a source.
- What do the prompt cache and batch settings do?
- The cache setting is the share of input tokens billed at the model’s cached input price, for prompts that start with the same text as earlier requests. The batch setting is the share of requests sent through an asynchronous batch API, which most providers price lower than instant requests. Both only lower the cost of the traffic they apply to.
- How is the GPU cost per million tokens calculated?
- Monthly rental cost divided by the tokens the GPUs produce in a month. Tokens come from a throughput figure (a published benchmark where one exists, or a figure you enter) multiplied by the busy share and the hours switched on. Benchmarks measure maximum throughput, so treat the result as a best case.
- Does the estimate include everything on my bill?
- No. It covers token and GPU-hour charges. It leaves out cache write and storage fees, tool, search, image and audio charges, embeddings, storage, networking and data transfer, and any discounts you have negotiated. The methodology page lists every simplification.
- Are my inputs sent anywhere?
- No. The calculator runs entirely in your browser, and the figures you type are not sent to us or stored.
Country pages
Write about AI or developer tools? Add this calculator to your site for free.