What are tokens in AI?

A token is a piece of a word, not a word

Models do not read letters and they do not read words. Before your message reaches the model, a tokenizer splits it into pieces from a fixed vocabulary of tens of thousands of fragments. Common words are one token. Rare words are two or three. "The" is one token; "tokenization" is typically two or three; a long URL can be dozens.

The rule of thumb OpenAI publishes for English text: one token is about four characters, or about three quarters of a word. So 100 tokens is roughly 75 words, and 1,000 tokens is a page and a half.

Two things make the count drift. Code runs denser, closer to one token per three characters, because of symbols and indentation. And languages that are less represented in the tokenizer's training data cost more: the same meaning in Thai, Hindi or Arabic can take two to three times the tokens of the English version, which matters if you're paying per token and writing in one of them.

You can see it for yourself. Paste any text into the token cost calculator and it shows the count and the price on each model.

Input tokens and output tokens are two different prices

Every message has two halves. Input tokens are everything the model reads: your question, the system instructions, the conversation so far, any file you attached. Output tokens are what the model writes back.

Providers charge the two at different rates, and the gap isn't small. The table is list price per million tokens from the AI model cost index, with prices fetched from OpenRouter on 2026-08-20.

Output is dearer because of how the model runs. Reading the input happens in one parallel pass over all the tokens at once. Writing the output happens one token at a time, each new token requiring another pass through the whole model. A 500-token answer is 500 sequential passes.

What a real message costs, worked through

Take a typical exchange: a 300-word question with a paragraph of context, answered in 400 words. That's about 400 input tokens and 530 output tokens.

• On GPT-5.6 Luna: 400 × $0.20 per million plus 530 × $1.20 per million, which is $0.00008 plus $0.00064, about $0.0007. Seven hundredths of a cent.
• On Claude Sonnet 5: 400 × $2 plus 530 × $10, per million, which is $0.0008 plus $0.0053, about $0.006. Just over half a cent.
• On GPT-5.5: 400 × $5 plus 530 × $30, per million, which is $0.002 plus $0.0159, about $0.018. Nearly two cents.

Notice that the output half is 80 to 90 percent of the bill in every row. The shape of the message, not just the model, sets the cost: a one-line question that asks for a 2,000-word essay is the expensive pattern, and a long document with a one-line question about it is the cheap one on the same model.

Zoom out and the spread across the whole catalogue is the real story. Every priced model in the AI model cost index sits on the line below by what one standard answer costs. The cheapest, Ling-3.0-flash, is $0.000053; the dearest, Claude Opus 4.7 in its fast mode, is $0.105: roughly 2,000 times apart for the same 1,500 tokens.

The other multiplier is history. In a chat, the model re-reads the whole conversation on every turn, so the input count grows with each message. A 40-turn conversation sends the first message 40 times. That's why long threads slow down and why what is a context window is the limit you eventually hit.

Why a subscription hides the meter but does not remove it

Consumer plans charge a flat monthly fee, and underneath that fee every provider is still counting tokens. The count surfaces as message caps, the "you have reached your limit" wall, and the quiet switch to a smaller model when the budget runs low.

Whizi meters the other way round. Each model has a fixed credit price per message: 1 credit for the house model, 10 for Claude Sonnet 5, 20 for Claude Opus 5 and GPT-5.5, and 100 at the very top of the ladder. The number never depends on how long your message or the answer was. The ladder is anchored to a real token turn, 3,000 tokens in and 800 out on the house model, which costs about $0.00156 in provider spend and defines one credit. How Whizi credits work has the full table.

The trade is deliberate. A per-token meter is exact and unpredictable; a per-message credit is slightly generous to long messages and slightly stingy to short ones, and you always know the price before you press send.

Tokens are also the capacity limit

The second reason tokens matter is capacity. Every model has a context window measured in tokens: the maximum it can consider at once, input and output combined. Claude Haiku 4.5 holds 200,000 tokens. GPT-5.6 Terra, Claude Sonnet 5 and Gemini 3.7 Flash hold about 1 million. At 750 words per 1,000 tokens, a million tokens is roughly 750,000 words, or about 1,500 pages.

The window is not free. Filling it costs input tokens on every turn, so a 300,000-token document asked ten questions is 3 million input tokens, which is $6 on Claude Sonnet 5 at list price. Retrieval, where the app searches the document and sends only the relevant pages, exists to avoid exactly that bill. What is RAG explains the mechanism, and context window comparison has every model's limit in one table. Send the pages the question needs, never the file.