LLM API Cost Calculator
What does the GPT, Claude or Gemini API cost? Enter requests and tokens and compare monthly costs in USD and euros, free and without sign-up.
Use case
Compare models
GPT-5 mini
OpenAI
Claude Haiku 4.5
Anthropic
Mistral Large
Mistral
GPT-5
OpenAI
Gemini 2.5 Pro
Claude Sonnet 5
Anthropic
Claude Opus 4.8
Anthropic
| Model | Provider | USD / month | EUR / month |
|---|---|---|---|
| GPT-5 mini Cheapest model | OpenAI | $8.00 | 7.04 € |
| Claude Haiku 4.5 | Anthropic | $23.00 | 20.24 € × 2.9 |
| Mistral Large | Mistral | $34.00 | 29.92 € × 4.3 |
| GPT-5 | OpenAI | $40.00 | 35.20 € × 5.0 |
| Gemini 2.5 Pro | $40.00 | 35.20 € × 5.0 | |
| Claude Sonnet 5 | Anthropic | $69.00 | 60.72 € × 8.6 |
| Claude Opus 4.8 | Anthropic | $115.00 | 101.20 € × 14.4 |
Providers' list prices as of 14 July 2026 · EUR conversion at a rate of ≈ 0.88 (as of 14 July 2026) · excluding caching or batch discounts · all figures without guarantee.
What are tokens?
A token is the smallest billing unit of a language model. In German, one token corresponds to roughly 3 to 4 characters, so an A4 page of text with around 3,000 characters comes to about 750 to 1,000 tokens.
What matters for the cost calculation: input tokens (your prompt including context) and output tokens (the model's answer) are counted separately and priced differently. Anyone who wants a realistic estimate of what an LLM application costs therefore has to keep both sides in view: not just the request itself, but also the expected length of the answer.
Provider pricing models
All major providers bill in US dollars per 1 million tokens, and for every model the output is considerably more expensive than the input: by a factor of 3 to 8 for the models compared here. There are three model classes to choose from: frontier models such as the Claude Opus class (5 USD input / 25 USD output), mid-range models such as the Claude Sonnet class (3 / 15 USD) or GPT-5.x (1.25 / 10 USD), and efficiency models such as Claude Haiku 4.5 (1 / 5 USD) or GPT-5 mini (0.25 / 2 USD).
The comparison calculator above shows all models side by side with their current prices, so you can see at a glance which model is the most economical for your specific use case.
Hidden cost factors: caching, batch and more
The pure token calculation is only half the story. Prompt caching discounts repeated parts of a prompt, sometimes considerably, and batch APIs typically offer around a 50% discount for asynchronous processing compared with real-time requests. Without these levers you often pay more than necessary.
Conversely, long system prompts and RAG contexts drive costs up, because they are billed again with every single request. Agent workflows with several tool calls multiply token consumption per user request on top of that, and retries after errors also show up on the bill.
Which model pays off when?
The basic rule: simple, high-volume tasks such as classification, FAQ chat or text extraction belong on an efficiency model, complex reasoning or demanding coding tasks on a frontier model. The difference is considerable: for an identical chatbot workload of 10,000 requests per month, Claude Haiku 4.5 costs around 23 US dollars, whereas the Claude Opus class costs around 115 US dollars, a factor of 5 for the same task.
If you do not route cleanly and run everything through an expensive model, you pay many times more without a corresponding gain in quality.
Cost optimisation in AI projects
Only optimise after you have measured; the calculator above gives you the baseline figures. The most effective levers, in order of impact: use a smaller model or model routing, use prompt caching for recurring content, shorten system prompts, limit output length, and use batch processing for everything that does not have to run in real time.
Planning an AI integration with these levers from the start often saves a double-digit percentage of the ongoing API costs. Experienced consultants can help you design such integrations cost-effectively.
Frequently asked questions
What does the ChatGPT API cost per month?
That depends on the number of requests and the token volume. Example: a chatbot with 10,000 requests per month (an average of 800 input and 300 output tokens per request) costs around 40 US dollars a month with GPT-5 (1.25 USD input / 10 USD output per 1 million tokens) and around 8 US dollars with GPT-5 mini. The formula: requests × (input tokens ÷ 1 million × input price + output tokens ÷ 1 million × output price).
How are the costs of an LLM API calculated?
All major providers bill per token, separately for input (your prompt including context) and output (the model’s answer). Output tokens are considerably more expensive than input tokens, by a factor of 3 to 8 for the common models. The monthly cost is: number of requests × (input tokens ÷ 1 million × input price + output tokens ÷ 1 million × output price). The calculator above runs exactly this calculation for all selected models at once.
What is a token and how many tokens does a text have?
A token is the smallest billing unit of a language model; in English a token corresponds to roughly 4 characters. An A4 page of text with around 3,000 characters is about 750 to 1,000 tokens. 1 million tokens, the usual pricing unit of the providers, therefore corresponds to roughly 1,000 to 1,300 pages of text.
Which LLM is the cheapest?
For simple, high-volume tasks, efficiency models are the cheapest: GPT-5 mini (0.25 USD input / 2 USD output per 1 million tokens) and Claude Haiku 4.5 (1 USD / 5 USD). Frontier models such as the Claude Opus class (5 USD / 25 USD) are only worthwhile for complex reasoning or demanding coding tasks. For the identical chatbot workload of 10,000 requests per month, the difference is 23 US dollars (Haiku 4.5) versus 115 US dollars (Opus class), a factor of 5.
How can I reduce my LLM API costs?
The five most effective levers: first, use a smaller model or model routing, sending simple subtasks to the cheap model and only complex ones to the expensive one. Second, use prompt caching: repeated prompt parts are billed with discounts of up to 90 percent depending on the provider. Third, use batch APIs for non-real-time jobs (typically around 50 percent discount). Fourth, shorten system prompts and attached context, because input tokens are charged on every request. Fifth, limit output length, because output tokens are the most expensive.
Provimedia
Let's talk about your project
Want to get more out of the result? We will show you how, with no obligation: a concrete quote instead of a sales pitch.
More free tools
AI Training Obligation Check
Are your employees required to receive AI training under the EU AI Act? 8 questions to find out.
AI Knowledge Test for Teams
10 questions, 3 minutes: how safely does your team use AI? With instant results and a level rating.
Compliance Quick Check
Which legal obligations apply to your company? 8 questions, instant results, no sign-up.