You have found a model you like and a price next to it: $2 per million tokens. That number is useless until you can answer one question — how many tokens does your app actually use?
Most people guess, and most people guess low. Not because tokens are complicated, but because almost everything in a request is made of tokens, and only a small part of it is the thing you typed.
Here is what a token is, how to count one, and where the surprises hide.
Prices last checked: August 26, 2026. Live figures are on the ElliSekiz compare page.
What a token actually is
AI models do not read letters or words. They read tokens — small pieces of text from a fixed vocabulary the model was built with.
A token can be a whole word, part of a word, a single character, or a piece of punctuation. Roughly:
cat→ one tokentokenization→ often three or four pieces, liketoken+iz+ation!→ one token- The space before a word usually travels with it, so
theis one token, not two
Why break text up this way? A model needs a fixed list of things it can recognise. A list of every English word would be enormous and still miss names, typos and other languages. A list of about 100,000 sub-word pieces covers nearly anything, because unfamiliar words can be spelled out of familiar fragments.
That is the whole idea. Text goes in, gets chopped into known pieces, and you are billed per piece.
How many tokens is your text?
OpenAI's rule of thumb for English is one token ≈ 4 characters ≈ ¾ of a word, so 100 tokens is about 75 words.
That is fine for a sanity check and bad for a budget, for three reasons.
Every model tokenises differently. The same sentence produces different counts on different models — sometimes very different. Anthropic's own documentation states that Claude 4.7 and later use a newer tokenizer where "the same input text produces approximately 30 percent more tokens than on earlier models." Same text, same vendor, 30% more tokens, purely because of a version change.
Other languages cost more. The vocabulary of most models is dominated by English. Text in languages that are less represented gets broken into more, smaller pieces — so the same meaning costs more tokens.
Code and structured text behave oddly. Indentation, brackets and long identifiers fragment in ways prose does not.
The reliable move is to count rather than estimate. Providers give you a way to do this: Anthropic has a token-counting endpoint that is free to call, and OpenAI publishes a tokenizer tool. Count once with your real prompts and you never have to guess again.
The part that surprises people: everything is tokens
Your question is a small share of what you pay for. Anthropic's documentation publishes the exact counts for a few minimal requests, and they are worth reading closely:
| What was sent | Input tokens |
|---|---|
| System prompt "You are a scientist" plus the message "Hello, Claude" | 14 |
| The question "What's the weather like in San Francisco?" plus one tool definition | 403 |
| One image plus "Describe this image" | 1,028 |
| One PDF plus "Please summarize this document." | 2,188 |
Read the second row again. The question is nine words. Describing a single tool — its name, its description, its input schema — costs roughly 390 tokens. An agent with fifteen tools pays that on every single request, before the user has typed anything.
The third and fourth rows are the same story for media. An image is about a thousand tokens. A PDF is a couple of thousand, and that was a short one.
Input tokens and output tokens are not the same price
Every text model publishes two rates. Output is always the expensive one.
| Model | Input $/1M | Output $/1M | Output costs |
|---|---|---|---|
| GPT-5 nano | $0.05 | $0.40 | 8× input |
| Gemini 2.5 Flash Lite | $0.10 | $0.40 | 4× input |
| GPT-5 mini | $0.25 | $2.00 | 8× input |
| Claude Haiku 4.5 | $1.00 | $5.00 | 5× input |
| GPT-5 | $1.25 | $10.00 | 8× input |
| Claude Sonnet 5 | $2.00 | $10.00 | 5× input |
| Claude Opus 5 | $5.00 | $25.00 | 5× input |
This has a direct practical consequence: telling the model to answer in three sentences instead of ten saves more money than shortening your prompt by the same amount. It is usually the cheapest optimisation available, and it takes one line of instruction.
What one conversation actually costs
A 10-turn chat. A 500-token system prompt, 100-token questions, 250-token answers.
The catch is that models have no memory between calls — the whole conversation is re-sent every turn. So turn 10 pays for turns 1 through 9 all over again.
- Tokens you typed: 10 × 100 = 1,000
- Input tokens you are billed for: 21,750
- Output tokens: 10 × 250 = 2,500
You typed a thousand tokens and paid for nearly twenty-two thousand.
| Model | Cost of that one conversation |
|---|---|
| GPT-5 nano | $0.002 |
| Claude Sonnet 5 | $0.069 |
Both are small. Multiply by 50,000 conversations a month and the difference is the difference between a rounding error and a real line item.
Four things that quietly inflate a token bill
- The system prompt, charged on every request. An 800-token system prompt across 50,000 messages is 40M tokens a month for text you wrote once. Prompt caching exists for exactly this — the provider stores the unchanging prefix and charges a fraction to re-read it.
- Conversation history. As shown above. Trimming or summarising old turns is the fix.
- Tool definitions. About 390 tokens each, on every request, whether or not the tool gets used.
- Images and PDFs. A thousand tokens and up, each.
None of these are hidden — they are all documented. They are just easy to leave out of an estimate.
Tokens are only one of several pricing units
Worth knowing before you generalise: most AI models are not priced in tokens at all. Of the 472 models tracked on ElliSekiz, 160 are priced per token — text and chat models. The other 338 are priced per image, per megapixel, per second of video, or per run.
So "how much does this model cost" has a different shape depending on what the model produces, and a per-token price can never be compared directly against a per-image one.
Conclusion
- A token is a small piece of text, not a word. Roughly ¾ of a word in English, and less in most other languages.
- Never budget on the rule of thumb. Count with the provider's own tool, against the exact model you plan to use.
- Your question is the small part. System prompts, conversation history, tool definitions and attachments are the large part.
- Output tokens cost 4× to 8× more than input tokens, so shorter answers are the cheapest saving.
Explore more on ElliSekiz
Browse the chat models. Input and output token prices for every text model, side by side, on the chat models page.
Then try one. Every model page has a playground — run your real prompt and read the actual token counts back before you multiply anything by 50,000.
Sources
- Anthropic — Token counting (documented example counts, tokenizer change, free counting endpoint)
- OpenAI Help Center — What are tokens and how to count them? (rule of thumb for English)
- OpenRouter models API,
https://openrouter.ai/api/v1/models(token prices, read 2026-08-26)