The unit a model reads and writes: a whole common word, part of a rarer one, a number, a space or a sign. Models are priced and limited in tokens, not words, and each maker cuts text into tokens its own way.

How many tokens is a word?

In plain English, about 1.3 tokens a word, so 100 words are about 130 tokens. Code, numbers and other languages take more. The example on this page, which mixes all of them, is 137 tokens for 95 words on OpenAI's tokenizer.

Why is the same text a different number of tokens on each model?

Each maker trains its own tokenizer on its own mix of text and languages. The example on this page is 137 tokens with OpenAI's tokenizer and 144 with Alibaba's, so a lower price per token can still mean a larger bill.

How do I count Claude tokens exactly?

Anthropic does not publish its tokenizer, so no page outside its API can count Claude tokens exactly. Its API has a free token counting endpoint for a message, and Anthropic itself calls the result an estimate. Here you get a range, from OpenAI's count to 1.75 times it, based on Anthropic's own note and an independent test.

How do I count Gemini tokens?

Google does not publish the Gemini tokenizer. Its API has a method that counts the tokens of a request, and its rule of thumb is about 4 characters a token, which is what this counter uses.

Is my text sent anywhere?

No. The tokenizers are downloaded to your browser and the text is counted there. Nothing you paste or open is sent to us or to any maker.

Token counter

Paste a prompt, a document or some code. See how many tokens it is on every current model, what it costs to send, and how much of each window it fills. The counting happens in your browser.

Prices read from each maker's page on 14 October 2026

Count a text Cost calculator Every price

The text stays in your browser. Nothing is sent anywhere.

Characters: 543
Words: 95
Lines: 9

Tokens

OpenAIAbout137
AnthropicRange137–240
GoogleAbout136
xAIAbout137
DeepSeekExact137
Moonshot AIExact140
AlibabaAbout144

How each maker is counted

What it costs to send

The answer you expect

words

At the output price, in as many tokens a word as your text takes.

The answer

136 to 240 tokens, by model. Sending it once with a 300-word answer costs 0.023 cents on GPT-6 Luna and 3.2 cents on Claude Fable 5.1.

0.023 centsGPT-6 Luna3.2 centsClaude Fable 5.1

Each of the newest models, cheapest first
Model, cheapest first	Tokens	To send	Answer	Window
GPT-6 Luna OpenAI	137	0.0014 cents	0.022 cents	1%
Qwen3.8-Flash Alibaba	144	0.0022 cents	0.021 cents	·
Claude Haiku 5.5 Anthropic	137–240	0.0014–0.0024 cents	0.022–0.038 cents	1%
DeepSeek V4.1 Flash DeepSeek	137	0.0041 cents	0.052 cents	1%
Gemini 3.5 Flash-Lite Google	136	0.0041 cents	0.11 cents	1%
Gemini 3.8 Flash Google	136	0.01 cents	0.16 cents	1%
DeepSeek V4 Pro DeepSeek	137	0.018 cents	0.17 cents	1%
Grok 4.7 xAI	137	0.027 cents	0.26 cents	1%
Qwen3.8-Max Alibaba	144	0.029 cents	0.27 cents	·
Gemini 4 Argon Google	136	0.027 cents	0.43 cents	·
GPT-6.1 Sol OpenAI	137	0.027 cents	0.43 cents	1%
Claude Sonnet 5.5 Anthropic	137–240	0.027–0.048 cents	0.43–0.76 cents	1%
Kimi K3 Moonshot AI	140	0.042 cents	0.66 cents	1%
Claude Opus 5.5 Anthropic	137–240	0.055–0.096 cents	0.87–1.5 cents	1%
GPT-6 Astra OpenAI	137	0.14 cents	2.2 cents	1%
Claude Fable 5.1 Anthropic	137–240	0.14–0.24 cents	2.2–3.8 cents	1%

Standard list prices. Window is the share of the model's context window the text and the answer fill together. A maker that charges more above a prompt size is priced at its standard rate here; the cost calculator shows where a higher one starts.

How each maker is counted

OpenAIGPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna: GPT-5.6 is counted with o200k_base, the tokenizer OpenAI publishes for it, so its count is exact. OpenAI has not published the tokenizer of GPT-6, so its count is the same o200k_base figure, marked as about. Source 1 Source 2
AnthropicClaude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5: Anthropic does not publish its tokenizer. It says Claude 4.7 and later produce about 30 percent more tokens for the same text than earlier models, and an independent test counted TypeScript at 1.73 times OpenAI's figure. The range runs from OpenAI's count to 1.75 times it. Source 1 Source 2
GoogleGemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite: Google does not publish the Gemini tokenizer. Its rule is that a token is about 4 characters, so the count is the characters divided by 4. Source
xAIGrok 4.7: xAI publishes neither its tokenizer nor a rule for estimating. The figure is OpenAI's o200k_base count, a rough guide only. Source
DeepSeekDeepSeek V4 Pro, DeepSeek V4.1 Flash: Counted with the tokenizer DeepSeek publishes with V4-Pro and V4.1 Flash, the same models its API serves. Source 1 Source 2
Moonshot AIKimi K3: Counted with the tokenizer Moonshot AI publishes with Kimi K3, the model its API serves. Source
AlibabaQwen3.8-Max, Qwen3.8-Flash: Counted with the tokenizer Alibaba publishes with its open Qwen3.8 models. Alibaba does not say its API models use the same one, so the count is marked as about. Source

How the counting works

A tokenizer cuts text into the pieces a model reads. It is trained on the maker's own mix of text, so a word that is one piece for one model can be three for another. That is why a price per token only compares models fairly once you know how many tokens your own text is on each.

OpenAI, DeepSeek and Moonshot AI publish the tokenizers of their models, and Alibaba publishes the one of its open Qwen models. The counter downloads those files to your browser and counts with them, the same way the makers' own tools do; we check every count against those tools. A count marked Exact matches what the maker bills.

Anthropic, Google and xAI do not publish theirs. For them the counter uses the maker's own rule where there is one: Google's 4 characters a token, and Anthropic's note that its newest models produce about 30 percent more tokens than before, shown as a range. xAI gives no rule, so its figure is only a guide. "How each maker is counted", above, gives the basis and sources of each.

Your text never leaves your browser. The answer length is priced in as many tokens a word as your own text takes, and a model whose window cannot hold the text and the answer together is marked.

Questions

What is a token?: The unit a model reads and writes: a whole common word, part of a rarer one, a number, a space or a sign. Models are priced and limited in tokens, not words, and each maker cuts text into tokens its own way.
How many tokens is a word?: In plain English, about 1.3 tokens a word, so 100 words are about 130 tokens. Code, numbers and other languages take more. The example on this page, which mixes all of them, is 137 tokens for 95 words on OpenAI's tokenizer.
Why is the same text a different number of tokens on each model?: Each maker trains its own tokenizer on its own mix of text and languages. The example on this page is 137 tokens with OpenAI's tokenizer and 144 with Alibaba's, so a lower price per token can still mean a larger bill.
How do I count Claude tokens exactly?: Anthropic does not publish its tokenizer, so no page outside its API can count Claude tokens exactly. Its API has a free token counting endpoint for a message, and Anthropic itself calls the result an estimate. Here you get a range, from OpenAI's count to 1.75 times it, based on Anthropic's own note and an independent test.
How do I count Gemini tokens?: Google does not publish the Gemini tokenizer. Its API has a method that counts the tokens of a request, and its rule of thumb is about 4 characters a token, which is what this counter uses.
Is my text sent anywhere?: No. The tokenizers are downloaded to your browser and the text is counted there. Nothing you paste or open is sent to us or to any maker.