Five free tools for anyone building with AI APIs: estimate token counts for GPT, Claude and Gemini models, work out what a request will cost, and check whether your text fits a model's context window. No sign-up, no API key — everything runs in this page, in your browser.
Paste or type any text. Counts update as you type, with estimated token totals for the three major tokenizer families.
These token numbers are estimates. TokenKit approximates each family with its average characters-per-token ratio (about 4.0 for GPT-style, 3.5 for Claude-style, 4.2 for Gemini-style English text). Real counts differ by text — code, numbers, URLs and non-English languages tokenize differently. For an exact count, use the provider's own tokenizer or the token-count endpoint in their API.
Pick a model, enter input and output tokens, and see exactly what one request costs — plus what 1,000 requests would run.
Standard on-demand rates, before prompt-caching discounts, Batch/Flex tiers or regional pricing. Reasoning models also bill hidden "thinking" tokens as output — budget extra if you use one.
| Model | Input / 1M | Output / 1M | Context window |
|---|
Prices checked October 2026 — always confirm on the provider's pricing page before budgeting. Rates are the standard API list prices published by OpenAI, Anthropic and Google.
Paste text and see how much of each model's context window it would fill. Bars turn amber past 80% and red when the text no longer fits.
Estimated tokens: 0 (GPT-style), 0 (Claude-style), 0 (Gemini-style). Remember the window also has to hold the model's reply — leave headroom.
One prompt, every model: the estimated input tokens and what sending that prompt once would cost on each model in the price table.
| Model | Est. input tokens | Input price / 1M | Cost per send ⇅ |
|---|
Input cost only — what you pay to send the prompt, before the model's reply. Token counts are estimated with each model's own tokenizer family, so the same text can show slightly different counts per row. Cheapest row is highlighted. Prices checked October 2026.
Type in any box and the others update. Handy for turning a word-count target ("write about 1,500 words") into a token budget, or the other way round.
These are the standard rules of thumb, not exact conversions. Real token counts vary by tokenizer and by content: code, tables, numbers and non-English text usually need more tokens per word. For anything you will be billed on, check with the Token Counter — or better, the provider's own tokenizer.
The unit every AI API bills you in — explained in two minutes.
Language models don't read text letter by letter, and they don't read whole words either. Before a model sees your text, a program called a tokenizer chops it into chunks called tokens. A token is usually a common word, part of a long word, a punctuation mark, or even a single space-plus-letter combination.
Take the sentence "Hello, world!" A GPT-style tokenizer splits it into 4 tokens:
Notice that the space before "world" belongs to the token — tokenizers treat " world" and "world" as different pieces. Long or unusual words get split into smaller pieces: "unhappiness" becomes un + happi + ness (3 tokens), and a rare word or a string of digits can take many more.
Three consequences matter when you build with these APIs:
A useful mental shortcut for English: 1 token ≈ ¾ of a word ≈ 4 characters. A page of text is roughly 650–700 tokens; a typical novel runs to about 100,000–130,000 tokens.
Straight answers about token estimates, prices and how this page works.
No — they are careful estimates. Exact counts require running the provider's own tokenizer (for example OpenAI's tiktoken library, Anthropic's token-counting API endpoint, or Google's countTokens API), because each tokenizer splits text according to its own learned vocabulary.
TokenKit uses each family's average characters-per-token ratio, which lands within a few percent for ordinary English prose. Expect bigger gaps on source code, JSON, URLs, text full of numbers, and languages other than English. Use these numbers for budgeting and comparing models, then confirm with the provider before a large production run.
Each provider trains its tokenizer on different data with a different vocabulary size. A tokenizer with a large vocabulary can store longer common chunks as single tokens; a smaller one splits the same word into more pieces. Newer tokenizer generations also change the balance — recent Claude models, for example, produce noticeably more tokens for the same text than earlier Claude tokenizers did.
That's why the same prompt can cost slightly different amounts on different models even at the same per-token price.
Often, and almost always downward for a given capability level. Providers cut prices when new model generations launch, run introductory pricing (two models in the table above are on dated introductory rates right now), and retire older models. The table on this page was checked in October 2026; treat it as a snapshot, and always confirm on the provider's own pricing page before committing a budget.
Input tokens are what you send: your prompt, system instructions, conversation history and any documents. Output tokens are what the model generates in reply. Output typically costs 4–5× the input rate because generating text is the expensive part — the model produces output one token at a time, while input is processed in parallel.
Watch out with reasoning models: their internal "thinking" tokens are billed as output too, even though you never see most of them.
No. Every tool on this page runs in JavaScript inside your browser tab. Nothing you paste is uploaded, stored or tracked — you can verify it by loading the page, disconnecting from the internet, and using the tools offline.
The biggest levers: use a smaller model for simple steps (classification, extraction, formatting) and save frontier models for hard reasoning; trim system prompts and conversation history instead of resending everything; cache repeated prompt prefixes where the provider offers prompt caching; cap output length; and use batch or flex tiers for work that doesn't need an instant answer. The Prompt Cost Comparison tool shows how wide the price gap between models can be for the exact same prompt.