Contents
At a glance
- Category
- Agent concept - model plumbing
- Rough English guide
- About 4 characters, or three quarters of a word, per token
- Tokenizes badly
- Code, JSON, IDs, phone numbers, non-English text
- Pricing
- Input and output billed separately, output usually several times dearer
- Agent effect
- Input tokens are re-billed on every step because the transcript is re-sent
- Discount lever
- Prompt caching on a stable prefix, where the provider supports it
What a token is
A token is the chunk of text a language model actually processes. Models do not read characters or words; a tokenizer splits text into a fixed vocabulary of fragments, and those fragments are what go in and come out. Common words are usually a single token, unusual ones split into several, and punctuation, digits and whitespace each take their share.
The frequently quoted approximation for English prose is about four characters, or three quarters of a word, per token. Treat that as a planning heuristic and nothing more. Code, JSON, product SKUs, phone numbers and non-English text all tokenize far less efficiently, which is why a JSON payload can cost several times what the same information costs as a sentence.
How tokens are billed
Providers price input and output separately, and output is normally several times more expensive. That asymmetry should shape your design: sending a long document and asking for a one-line classification is cheap; sending one line and asking for a long report is not. Most business automations are input-heavy, which is good news for the bill.
The subtlety in an agent is that input tokens are paid repeatedly. Every step re-sends the entire transcript, so the tokens from step one are billed again at steps two through eight. A ten-step agent does not cost ten model calls' worth of input - it costs closer to the sum of a growing series. Prompt caching mitigates this for the stable prefix at a discounted rate, which is why keeping the system prompt stable and at the top is worth real money.
Practical ways to spend fewer
Four things move the needle. Trim tool results to the fields you need before they enter the context. Ask for terse structured output rather than prose you then parse. Route easy work to a small model and reserve the expensive one for the step that needs it. And cap agent steps, because runaway loops are how token bills become surprises.
Also watch the units when comparing providers. Published prices are per million tokens, and a model that looks twice as expensive may use a tokenizer that produces fewer tokens for your text, or may finish the job in one call where the cheaper one needs three. Compare on cost per completed task using your own eval set, not on the headline rate.
What you should know
Tokens are also the reason certain tasks fail in ways that look like stupidity. Counting the letters in a word, reversing a string or doing precise arithmetic are all hard for a model that sees fragments rather than characters. Do not automate those with a prompt - give the model a tool that does the calculation and let it read the answer.
For budgeting a real automation, our AI automation cost page covers what the model line actually contributes to a monthly bill, which for most small-business builds is a smaller share than people expect next to platform fees and integration work.
What a token actually costs, in August 2026
A token is only meaningful once it has a price on it, so here are current published rates rather than a general statement. Per million tokens, input then output: gpt-5.6-sol $5.00 and $30.00; gpt-5.6-terra $2.00 and $12.00; gpt-5.6-luna $0.20 and $1.20; gemini-2.5-flash-lite $0.10 and $0.40; deepseek-v4-flash $0.22 to $0.44 and $0.66 to $1.32 depending on time of day. All checked against the vendors' own pricing pages in August 2026, and all liable to change.
Two structural facts matter more than the individual numbers. Output tokens cost several times input tokens on every one of those lineups, so a verbose model is more expensive than a long prompt. And cached input is a fraction of standard input - a tenth on OpenAI, around a thirtieth on DeepSeek - which makes a byte-stable request prefix the single largest lever available on any repeated call.
Token FAQ
How many words is a token?
In English prose, roughly three quarters of a word, or about four characters. It is only an approximation: common words are one token, rare words split into several, and code, identifiers, numbers and other languages are considerably less efficient.
Why is output more expensive than input?
Generating tokens requires a forward pass per token, while input can be processed in parallel. Providers price that difference through, which is why classification and extraction jobs are cheap and long-report generation is not. Design prompts to be input-heavy and output-light where you can.
Why does my agent's token usage grow so fast?
Because each step re-sends the whole conversation, including every previous tool result. If your tools return full API responses instead of the fields you need, the transcript balloons and every subsequent step pays for it again. Trim results at the tool boundary.
How do I estimate token cost before building?
Take a real example input, run it through your provider's tokenizer to get an input count, estimate output length, multiply by the published per-million rates, then multiply by expected runs per month and by the number of model calls per run. Add a healthy margin for retries.
Is the cheapest model per token the cheapest overall?
Not necessarily. A cheaper model that needs three attempts, or a longer prompt to behave, can cost more per completed task than a dearer one that gets it right first time. Compare on cost per successful run measured on your own eval set.
Why is the model bad at counting letters or doing maths?
Because it never sees individual characters or digits - it sees tokens. Character counting, string reversal and precise arithmetic are structurally awkward for it. Hand those to a tool that computes the answer and let the model read the result rather than prompting harder.
Primary source
Hugging Face - tokenizer documentation
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what token means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
