Most AI tools bill by usage, and usage is counted in tokens, not in words or pages. That detail is why the language of your prompts can affect your bill.
This article explains what tokens are, why language can change how many you use, and what is and is not proven about English being cheaper. You will also get a simple method to measure the difference with your own work before you change any habits.
How AI billing works: tokens, not words
AI language models do not read text the way people do. Before a model processes your message, a tokenizer cuts the text into small pieces called tokens. A token can be a whole word, part of a word, or a punctuation mark.
Processing more tokens generally increases costs. That applies to what you send (your prompt, pasted documents, earlier conversation) and to what the model writes back. Even if you never see a bill directly, token counts can affect usage limits and how quickly a team burns through a shared budget.
So the practical question is simple: the same idea can take more or fewer tokens depending on how it is written, and the language is one of the things that changes this.
Why language can change the token count
Tokenizers split text into subword pieces. How many pieces a given word becomes depends on the tokenizer and its vocabulary. A word that is in the vocabulary as a single piece costs one token. A word that is not may be broken into several.
This means two things for you:
- The same sentence can use a different number of tokens in different languages.
- The same sentence can also use a different number of tokens in different AI models, because each has its own tokenizer.
What makes English straightforward to process
Tokenization is fairly simple for English because words are usually separated by spaces. The tokenizer has clear boundaries to start from. Languages such as Chinese, Japanese and Thai are different: they do not normally put spaces between words, so word segmentation is a significant task of its own.
That is a useful clue about why languages are handled differently. It is not, by itself, proof of lower prices. Whether a given text ends up with fewer tokens depends on the specific tokenizer, and the only reliable way to know is to count.
What we can and cannot say about savings
It is worth being precise here, because claims about AI costs are often repeated without evidence.
- Known: costs rise as the number of tokens processed rises.
- Known: tokenizers split text into subword pieces, and the number of tokens for a word varies by tokenizer and vocabulary.
- Known: English is simple to split into words because of spaces, unlike languages that need word segmentation.
- Not established here: that English always uses fewer tokens, or by how much. The size of any saving depends on the model, the text and the language pair.
So the honest position is this: English is a reasonable default to test first, and for many setups it may reduce token use, but you should confirm it with your own prompts instead of assuming a fixed discount.
How to test the savings yourself
You do not need special skills to check this. Most AI providers offer a token counter or show token usage in their dashboard or responses. Use whichever one belongs to the model you actually pay for.
- Pick three to five real prompts you use regularly, such as a meeting summary request, a customer email draft or a data-cleaning instruction.
- Write each prompt in English and in your usual language, saying the same thing in both.
- Count the input tokens for each version with the provider's tokenizer tool or usage report.
- Compare the outputs too. Ask for a reply in each language and compare the output token counts, since output is also billed.
- Check quality. A cheaper prompt that gives a worse answer can cost more once you retry and fix mistakes.
- Repeat with another model if you use more than one, because tokenizers differ.
If English comes out noticeably lower across your typical tasks, you have a measured reason to switch. If the difference is small, keep working in the language that is easiest and least error-prone for you.
Practical ways to apply English without losing quality
You do not have to give up your own language for everything. A few targeted habits tend to make sense:
- Write system instructions in English. Standing instructions are sent again and again, so any saving in them repeats with every request.
- Keep reusable templates in English. A prompt template used by a whole team multiplies its cost by the number of runs.
- Ask for English output when the result is for your own use. Notes, drafts and analysis you will only read internally do not need to be in another language.
- Translate at the end. If the final text must be in another language, work in English first and request a translation of the finished version only.
When English is not the right choice
There are clear cases where staying in your own language is better:
- The text depends on local wording, legal terms or cultural nuance that could be lost in translation.
- You would need to pay for or spend time on translating source material first, which can cost more than the tokens you save.
- Your English is not strong enough to state the task precisely. A vague prompt leads to retries, and retries cost tokens.
Other ways to cut token use
Language is only one lever. Because cost follows token count, anything that shortens what you send and receive also helps:
- Remove unneeded background and paste only the relevant part of a document.
- State the format and length you want, so the model does not write more than you need.
- Start a new conversation when the old one is no longer relevant, instead of carrying a long history.
- Reuse a tested prompt instead of experimenting from scratch each time.
These habits often matter as much as the choice of language, and they work in every language.
Conclusion
AI costs follow token counts, and token counts depend on the tokenizer and on how your text is split. English is simple to split into words, which makes it a sensible language to try first for prompts, templates and internal work. The available evidence does not, however, promise a specific saving.
The best approach is to measure. Count tokens for a handful of your real tasks in English and in your own language, compare the quality of the answers, and switch only where the numbers and the results both favor it.
