How Many Tokens Is My Prompt? A Practical Guide to Counting Tokens

Every AI model reads your text as tokens, not words — and tokens are the unit that decides both what fits in a context window and what a request costs. Yet almost nobody checks the count before pressing send. This guide explains what tokens actually are, the rules of thumb you can do in your head, why every provider counts the same text differently, and how to estimate a prompt’s token count in seconds with the free token counter.

What a token actually is

A tokenizer splits text into small pieces the model can digest. Common English words often survive as a single token — “hello”, “world”, ” pricing”. Rarer words get chopped into subword fragments: “tokenization” might become two or three pieces, and unusual names or misspellings shatter further. Punctuation, spaces, and line breaks are tokens too — they ride along attached to the words that follow them.

Two approximations from the model providers themselves hold up well for everyday English:

• One token is about 4 characters (spaces included).
• One token is about three-quarters of a word — so 100 words is roughly 130 tokens.

Neither rule is exact, because tokenizers are learned vocabularies, not arithmetic. But together they let you estimate almost anything without a calculator.

Why every provider counts differently

OpenAI, Anthropic, Google, and Meta each trained their tokenizer on their own corpus with their own vocabulary size. The same sentence can produce 21 tokens on one platform and 24 on another. Three factors drive the spread:

1. Vocabulary coverage. A tokenizer that has seen lots of code splits code more cheaply; one trained heavily on English wastes tokens on other languages.
2. Vocabulary size. Larger vocabularies merge more text into fewer tokens. Model generations ship new tokenizers, so counts shift between versions of the same product.
3. Script and content type. Chinese, Japanese, and Korean characters often cost a token each or more. Emoji are notorious multi-token spenders. Code with heavy punctuation runs more expensive than prose of the same length.

This is why any honest tool shows you a range, not a single number. Our token counter applies each provider’s published characters-per-token average separately — OpenAI, Claude, Gemini, and Llama — and reports the four estimates side by side with a range and an average, so you can budget against the worst case.

Why the count matters: two budgets, one meter

The context window is a hard wall. Every model has a maximum context — the sum of the system prompt, the conversation history, any attached documents, and the model’s own reply. Exceed it and the request fails outright. A prompt that fit yesterday can break after you paste one long document into it.

The meter is a soft drain. APIs bill per token in and per token out. A sloppy prompt repeated thousands of times in production is a real invoice line. And the reply counts too — an under-budgeted prompt quietly doubles its own cost once the model’s answer is added.

The practical discipline: know your numbers before you send.

Estimating tokens before you send

A real request carries more than the question you typed:

• the system prompt, which repeats with every call
• few-shot examples you included to steer behavior
• retrieved documents or pasted context
• the reply you expect, which you should reserve room for

Estimate each piece separately, add them, and keep the total under the limit with margin to spare. A few conversions worth memorizing:

Text	Approximate tokens
100 words	~130–170
1,000 words	~1,300–1,700
A tweet (280 chars)	~70
A typical page of prose (~3,000 chars)	~750

For the raw length check, run your draft through the word counter — then convert with the rule of thumb above, or let the token counter do both at once and show the per-provider spread.

The honest limits of any estimate

No browser tool can give you the exact number your provider will report, because the exact figure comes from the provider’s own tokenizer — and they change it between model versions. A good estimator gets within roughly 10–20% for everyday English prose and tells you plainly where it is less reliable: short snippets, dense code, emoji, and non-English text.

Use the estimate for planning — sizing a prompt, deciding what to cut, comparing whether a longer context is worth it. For billing and hard context-window decisions, verify with the usage figures your provider’s dashboard reports.

Count your own tokens

Paste any text into the free token counter and see OpenAI, Claude, Gemini, and Llama estimates side by side, instantly, entirely in your browser. Nothing is uploaded, no account is needed, and no AI model is involved — it is pure arithmetic on your device. For deeper analysis of what is making your text long, text statistics breaks down sentence length, vocabulary, and structure.