What is an AI token?
An AI token is the small chunk of text an AI model reads and writes in, and the unit it uses to measure and bill your usage. A token is roughly three quarters of an English word, so 1,000 tokens is about 750 words, or a dense page of text.
A helpful way to picture it: if words are the sentences you speak, tokens are the syllables the AI actually hears. It breaks everything down into these small pieces first, because that is the only form it can work with.
When you type a question to an AI tool, your words are broken into tokens before the model can process them. Its answer is also produced token by token. Everything the AI does, it does in tokens.
Why does AI use tokens instead of words?
AI uses tokens because they let a model handle any language, spelling, or symbol consistently, not just tidy English words. A token is often a sub-word fragment, which is why the count rarely matches the number of words you wrote.
A simple example makes this clear. A common word like "the" is a single token. A longer word like "monetization" may be split into three tokens, such as "mon", "etiz", and "ation". Numbers, punctuation, emoji, and Chinese characters each carry their own token cost too.
This design has a real benefit: it means one AI model can handle English, Chinese, code, and even emoji without needing a separate system for each. The trade-off is that token counts feel unintuitive, because they follow the fragments the model sees, not the words you recognise.
For Chinese text the ratio is different again. Chinese characters often use more tokens per character than English does per word, so a Chinese prompt of the same visible length can cost more to process. This matters for Hong Kong businesses working in mixed Chinese and English.
How does token-based pricing work?
Token-based pricing means you pay for what you actually use, measured in tokens, rather than a flat monthly fee. The bill is calculated as the number of input tokens times the input rate, plus the number of output tokens times the output rate.
There are two prices to watch, and they are not the same:
Input tokens are everything you send in: your question, plus any documents or instructions you attach.
Output tokens are everything the AI writes back to you.
Output tokens usually cost three to eight times more than input tokens. The reason is mechanical: the model reads your whole prompt in one fast pass, but has to write its answer one token at a time, which takes far more computing power.
How much do AI tokens actually cost?
Token prices are quoted per million tokens and vary widely by model. As of 2026, a cheap, fast model such as Gemini Flash Lite runs around US$0.08 per million input tokens and US$0.30 per million output tokens, while a top-tier model like Claude Opus 4.5 costs about US$5 input and US$25 output per million.
Those numbers sound abstract, so translate them. One million tokens is roughly 750,000 words, about nine average novels. For most small businesses, everyday tasks like drafting emails or summarising documents cost a few cents each on a cheaper model.
It also helps to know that prices have fallen sharply and keep moving. The cheapest capable models in 2026 cost a small fraction of what similar quality cost a year earlier, which is a large part of why AI is now affordable for an ordinary Hong Kong SME rather than only big firms with big budgets.
The practical lesson is that the model you choose matters more than how carefully you type. Using a flagship model for a simple task can cost fifty times more than a light model that would have done the job just as well.
What does token cost look like in a real task?
A worked example makes tokens concrete. Say a Hong Kong retailer asks an AI to summarise a two-page supplier email into three bullet points, and repeats this fifty times a month.
The two-page email is roughly 1,200 input tokens, and a three-bullet summary is around 150 output tokens. On a cheap model at US$0.08 input and US$0.30 output per million, each summary costs about 0.014 US cents. Fifty of them cost under one US cent for the whole month.
Now run the same task on a flagship model at US$5 input and US$25 output per million. Each summary jumps to roughly 1 US cent, and the month costs around 40 US cents. Still tiny, but almost fifty times more, for a job the cheap model handled perfectly. Multiply that gap across thousands of tasks and it becomes a real line on your budget.
The point is not the exact figures, which change often. It is the shape: everyday AI work is astonishingly cheap on the right model, and needlessly expensive on the wrong one.
Why does this matter for a small business?
It matters because token pricing is why two businesses using "the same AI" can get wildly different bills. Understanding tokens turns AI cost from a mystery into something you can predict and control.
Consider a shop that uses AI to answer customer enquiries. If each reply includes a long block of background instructions sent every single time, those input tokens repeat on every message and quietly inflate the bill. Trimming that context, or reusing it smartly, can cut the cost sharply.
The flip side is opportunity. Once you understand that a summary costs a fraction of a cent, you stop rationing AI and start using it freely for the small tasks that add up, freeing hours of staff time for a trivial cost.
How can you keep your AI token costs down?
You keep token costs down by matching the model to the task, keeping prompts lean, and reusing repeated context instead of resending it. Small habits compound into meaningful savings at scale.
Four practical moves:
None of this requires technical skill. It is closer to switching off lights you are not using: small, sensible habits that quietly protect your margins while you get on with running the business.
1. Use a lighter model for simple work. Reserve the expensive flagship models for genuinely hard tasks. Most everyday drafting and summarising runs perfectly well on a cheaper model.
2. Keep prompts focused. Long, rambling instructions are all input tokens you pay for. Say what you need clearly and stop.
3. Ask for the length you need. If you want a three-line summary, say so. Letting the AI write ten paragraphs you did not want is paying for output you will delete.
4. Reuse context with caching. If you send the same long background on every request, most providers offer prompt caching that discounts repeated input by 75% to 90%. Ask your tool or provider whether it is enabled.
Common misconceptions about tokens
The most common misconception is that a token equals a word. It does not; it is usually a fragment, so your token count is almost always higher than your word count.
A second myth is that only your question costs money. In fact the AI's answer usually costs several times more per token than your question did, so long-winded replies are the bigger expense.
A third is that a monthly subscription and token pricing are the same thing. Consumer subscriptions bundle a usage allowance into a flat fee; token pricing, used by business tools and APIs, charges for exactly what you consume. Many businesses use both, for different purposes.
Frequently asked questions
Do Chinese characters cost more tokens than English?
Often, yes. Chinese text tends to use more tokens relative to its visible length than English, so a bilingual Hong Kong workload can cost a little more to process than an English-only one of the same apparent size.
How can I see how many tokens I am using?
Most business AI platforms show token usage in a dashboard, and free online token counters let you paste text to estimate a prompt before you send it.
Should I worry about tokens if I only use a chatbot subscription?
Less so. A fixed monthly plan hides the token maths behind a flat fee. Tokens matter most once you build AI into your own tools or use a pay-as-you-go business service.
Does a longer conversation cost more?
Yes. Many AI tools resend the earlier parts of a chat as input on each new message so the model remembers the thread. A long conversation therefore costs more per message than a short one, because the history is billed again each time.
The takeaway
Tokens are the hidden meter running behind every AI tool. Once you understand that you pay per fragment of text, that answers cost more than questions, and that the model you pick sets the rate, AI pricing stops being a black box.
You do not need to count tokens by hand. You just need to know they exist, so you can choose the right tool and use it without fear of a surprise bill. We understand AI. UD stands with you.
Want AI that fits your budget?
Understanding tokens is the first step to controlling AI costs. The next is choosing the right tools and setting them up so they deliver value without waste. UD has helped Hong Kong businesses do exactly this for 28 years, and we will walk you through it step by step, from your first estimate to a setup that pays for itself.
