What will your AI cost?
See what an AI feature will cost each month on every major model. Pick what you are building and how much it will be used, or paste your own prompt to count its tokens. It runs in your browser, needs no API key, and nothing you type leaves this page.
Your project
Pick one to see what switching would save or cost.
What it costs
Working out the cost on every model.
Prices checked October 20, 2026. Email me when a price changes for my models Planned for Pro
How we calculate this
Fine-tune the estimate
The scenario filled in typical sizes. Change any of them, or paste your own prompt for a count on every model.
Prompt
ExampleStanding instructions are the system prompt: the same text sent before every call. The new text is what changes each time, such as a user message, a document or a tool result. These counts are used for every model as entered; the same text is a different number of tokens on each model, so paste the text when you have it to get an estimate per model.
Answers and discounts
When many calls start with the same instructions, providers can reuse them from a cache and charge a fraction of the input price for that part.
How prompt caching works
Prompt caching: the provider keeps the processed start of your prompt for a few minutes. If the next call begins with exactly the same text, those tokens are read from cache at the cache-read price instead of the full input price. Anything after the first changed character is billed normally, so put fixed text first and changing text last.
A busy app that sends the same instructions many times a minute can get close to 100%. Calls that are minutes apart, or instructions that change per user, get much less. Some providers cache automatically; others need you to mark the part to cache, and Anthropic bills the first write to the cache at 1.25 times the input price (2 times for a one-hour cache), which this calculator does not add.
Batch processing is about half price at Anthropic, OpenAI, Google and Mistral when answers can wait up to a day. It suits back-office work such as sorting a backlog overnight, not live chat. Check your provider's batch terms.
Batch answers can take hours, so batch prices do not suit live chat. Use them for work that can wait.
Counts
estimatesCost on every model
Context fit is the share of each model's context window that the largest call fills.
| Compared with yours |
|---|
Prices are USD per 1M tokens at standard rates from each vendor's pricing page.
Shared budgets and saved scenarios for your team Planned for Team
Monthly cost
Where the tokens go
Tokens, prices and the arithmetic
Everything on this page is computed in your browser from each vendor's published prices. Here is what the numbers mean and how they are put together.
What is a token?
Language models read and write text as tokens: whole short words, pieces of longer words, numbers, punctuation and spaces. A common English word is often a single token; a rare or long word is split into several. Providers bill by the token, separately for what you send (input) and what the model writes (output).
Rules of thumb
- In English, one token is roughly 4 characters or about 0.75 words, so 1,000 words come to roughly 1,300 to 1,400 tokens.
- Code, JSON, numbers, URLs and IDs use more tokens per character than prose.
- Many other languages, especially those not written in the Latin alphabet, need more tokens for the same meaning.
- These are approximations. The only exact count comes from the model's own tokenizer.
Why the same text is a different number of tokens on each model
Each model family has its own tokenizer, a fixed vocabulary of text pieces chosen from its training data. A tokenizer with more pieces for code or Korean needs fewer tokens for that text. Claude's tokenizer produces more tokens than OpenAI's for the same English text, so counting a Claude prompt with an OpenAI-style counter undercounts it. Even one vendor can change tokenizers between model generations.
This page estimates each family separately and marks those numbers with ≈. They are typically within 10 to 15% for English prose and looser for code and other scripts. If you add an Anthropic or Google AI Studio key in Settings, "Count exactly" asks the provider's own token-counting endpoint, which does not run the model.
How the price is calculated
Prices are quoted in dollars per million tokens. Input is everything you send: system prompt, conversation history, documents and tool definitions. Output is everything the model writes, including reasoning tokens on models that bill thinking as output.
Prices are quoted per million tokens, so each product is divided by 1,000,000. Batch processing halves every price where a provider offers it.
Output usually costs several times more than input: between 2 and 8 times for the current models in the table. Some models charge more for very long prompts, and some change price on a set date or by time of day; those cases are flagged in the table's notes, and the long-prompt rates are applied when your prompt crosses the threshold.
Why long conversations cost more per message
A model does not remember a conversation between calls. Each new message is sent together with the instructions and everything said so far, so the fifth message in a chat reads far more text than the first. The calculator adds up every call in a conversation this way. With "messages per conversation" set to 4, one conversation reads the instructions four times and the earlier messages and answers again and again, which is why a long chat costs more than four short ones.
How prompt caching saves money
Many calls start with the same long prefix: a system prompt, instructions, examples or a reference document. With prompt caching, the provider keeps that processed prefix for a few minutes, and the next call that begins with exactly the same text reads it at the cache-read price, between 2% and 50% of the normal input price for the models here. It only applies to an identical prefix, so keep fixed text at the start and put anything that changes (the user's question, today's date) after it.
The cache slider sets how much of the standing instructions is read from cache across all your calls. The rest is billed at the full input price, and so is the conversation history, which makes the estimate slightly cautious for long chats. Models without a published cache price are charged in full.
Worked example
A support bot has a 2,000-token system prompt, each question adds 500 tokens and each answer is about 300 tokens. It runs 10,000 times a day on Claude Haiku 5.5, at $0.10 per million input tokens, $0.50 per million output tokens and $0.01 per million cached tokens.
- Without caching: 2,500 input tokens × $0.10 + 300 output tokens × $0.50, divided by a million, is $0.0004 per call, $4 a day and $120 a month.
- With 90% of the system prompt cached: 1,800 cached tokens × $0.01 + 700 uncached tokens × $0.10 + 300 output tokens × $0.50 is $0.000238 per call, $2.38 a day and $71.40 a month.
Questions
How do I estimate what a chatbot will cost a month?
Pick "Support chatbot" at the top, enter how many conversations you expect a month and how many messages a typical conversation has. The summary shows the monthly cost on the cheapest model that fits and the most expensive one, and the table lists every model. If you already use a model, choose it under "The model you use now" to see what switching would save.
What can a fixed monthly budget buy?
Switch to "What a budget buys" and enter the budget. Each model then shows how many conversations or tasks that budget covers a month, most first.
How many tokens is 1,000 words?
About 1,300 to 1,400 tokens of ordinary English with an OpenAI-style tokenizer, and somewhat more on Claude. Technical writing, code and other languages run higher. Paste the text above to see an estimate for each model family.
Why is output more expensive than input?
The model reads all your input tokens in one parallel pass, but it writes output one token at a time, and every new token needs another pass through the model. Output takes far more computing time per token, and prices reflect that.
Are these prices current?
Direct provider prices were checked against each vendor's pricing page on October 20, 2026. OpenRouter prices load live from OpenRouter's public model list when you open this page, and are kept for up to 12 hours. Prices change often, so confirm on the vendor's page before you commit to a budget.
Do these include batch discounts or free tiers?
Batch prices are an option under "Answers and discounts": about half the standard price at Anthropic, OpenAI, Google and Mistral, for requests that can wait up to a day. Terms differ by provider and model, so check your provider's batch terms. Free tiers come with rate limits and are not counted; models with a free tier are tagged.
Is my prompt sent anywhere?
No. Counting and pricing run in this browser. The page downloads OpenRouter's public price list, which contains nothing from you. If you use "Count exactly", your text goes straight from your browser to Anthropic or Google with your own key, and nowhere else. Share links keep the prompt in the part of the address after #, which browsers never send to a server; the budget note never includes your prompt.