Vietnamese on AI Models: The Cost Now Depends on the Vendor

A short note for Vietnamese business owners and managers who are choosing an AI model for a chatbot or assistant, and for anyone who has heard that Vietnamese costs twice as much as English.

About this note. The numbers below were measured on 19 October 2026 with each vendor’s own token counter, on two public sets of the same sentences translated into several languages. The code and data are published so anyone can check or repeat the measurement. Where the note interprets the numbers, it says so.

Related work. Recent measurements by others already point the same way (see “Do other measurements agree?” below). This note adds a larger sample of real sentences, all current price tiers of the main vendors, a comparison of chat messages with formal text, and a comparison of models with the same list price.

Scope. This note compares token counts and list prices. It does not compare answer quality, and it does not estimate a full monthly bill (system instructions, conversation history and replies all add to real costs). Prices change often; all prices are as of 19 October 2026.

Why would the same message cost more in Vietnamese?

AI models do not read words. They read tokens, and vendors charge per token.

Each vendor splits text into tokens with its own tokenizer. Tokenizers learn their pieces mostly from English text, so English words often become one token each, while words in other languages can be split into several. The same message can therefore need a different number of tokens, depending on its language.

In 2023, this effect was large for Vietnamese. A 2023 study by researchers at the University of Oxford measured Vietnamese at about 2.45 times the tokens of English on the tokenizer used by GPT-4. Verified A rule of thumb is still repeated in Vietnamese developer writing in 2026: “chúng ta thường tốn gấp đôi token so với tiếng Anh cho cùng một lượng nội dung” Verified (in English: we usually spend twice as many tokens as English for the same content).

Tokenizers have changed since 2023. This note checks how far the rule still holds.

How much more does Vietnamese cost on today’s models?

On most current models, noticeably less than twice. The heatmap below shows the premium for short chat messages — the kind of text a customer sends to a shop’s chatbot — on the main models measured.

So the twice-as-much rule now describes one vendor rather than the market.

Three more points from the same measurement:

• The premium is the same across a model family’s price tiers. Every model within each family produced exactly the same token count on every sentence checked (200 per language in each dataset). Verified Choosing Haiku instead of Opus, or GPT-6 Luna instead of Astra, lowers the price per token but not the Vietnamese premium.
• Vietnamese has the highest premium of the languages compared. On chat messages, Vietnamese had a higher premium than Chinese, Russian or German on every model measured. Verified The heatmap shows the other languages for context.
• Formal text gives a similar picture. On a second, more formal set of sentences, the premiums were 1.30 times on Qwen, 1.35 to 1.36 times on Gemini, 1.48 times on GPT-6 and 1.84 times on Claude. Verified

Why can two models with the same list price cost different amounts for Vietnamese?

Because the price list shows the price per token, not how many tokens your text needs. Three models in the middle price tier all list the same input price, $2 per million tokens. For the same Vietnamese chat messages, their input costs were:

Prices as of 19 October 2026. Google has announced that Gemini 4 Argon’s input price rises to $4 per million tokens on 1 December 2026. Verified At that price its relative input cost would be about 200: still below Claude Sonnet 5.5’s 221, but no longer far from it.

At the same list price, the input cost for the same Vietnamese messages was about 2.2 times as high on Claude Sonnet 5.5 as on Gemini 4 Argon. Verified The gap is larger than the premiums alone suggest. Claude’s current tokenizer also uses more tokens for English than the other vendors’ tokenizers: about 10 tokens for an average English chat message, compared with about 7 on GPT-6. Verified So Claude starts from a higher base before the Vietnamese premium is added. This is a property of the tokenizer, not a judgement on how well a model handles Vietnamese.

One caution applies: the comparison covers the messages sent to the model. Replies are counted with the same tokenizer, and until Google’s 1 December change all three models also share the same output price ($10 per million tokens), so a similar gap is likely for replies. Plausible ^{8 9} Replies were not measured here.

What does this mean for a business choosing an AI model?

The premium is one factor in the choice, not the choice itself. For scale: within one vendor, the price per token of the cheapest and the most expensive model differs about 100 times (GPT-6 Luna at $0.10 against GPT-6 Astra at $10 per million input tokens), which is much more than any Vietnamese premium. A few practical points follow from the numbers:

1. Check quality first. A lower token count does not make a model answer better in Vietnamese. Shortlist the models that handle your customers’ questions well.
2. Compare the cost of your Vietnamese text, not only the price list. With the same list price, one model cost about 2.2 times as much as another for the same Vietnamese messages (at prices as of 19 October 2026).
3. Count a sample of your own messages. Anthropic, OpenAI and Google all provide token-counting tools. A few dozen real customer messages, counted in Vietnamese and in English on each shortlisted model, show the premium for your own messages.
4. Remember that prices change. The numbers here are a snapshot from 19 October 2026. A price change, or a new tokenizer, can change the comparison within months.

Do other measurements agree?

Yes, broadly. Independent measurements published in 2026 point in the same direction, using different texts and methods:

• A dataset from TokenSave, measured on 17 October 2026 with one customer-service prompt in 41 languages, found that Vietnamese needed about 26% more tokens than English on Gemini, 35% more on OpenAI’s current tokenizer and 87% more on Claude 5.5 (1.26, 1.35 and 1.87 times). Verified These are close to the 1.27, 1.40 and 2.00 times measured here.
• A benchmark published on Viblo in September, using one pair of paragraphs, found Vietnamese at 1.34 times English on OpenAI’s current tokenizer and 1.26 times on Qwen 3. Verified

The exact numbers differ by text, as expected. Where the sources measured the same vendors, the order is the same: Gemini lowest, OpenAI in the middle, Claude highest. Qwen, measured only here and on Viblo, was also among the lowest.

How was this measured?

The measurement used two public datasets in which the same sentences appear in many languages: MASSIVE ⁶, about 2,000 short virtual-assistant messages per language, and FLORES+ ⁷, about 2,000 formal sentences per language. For each model, the full set was counted in English and in Vietnamese (plus Chinese, Russian and German). The premium is the Vietnamese total divided by the English total.

Counts come from each vendor’s own token counter: Anthropic’s and OpenAI’s token-counting endpoints, Google’s Gemini API, and the published tokenizer for Qwen. The three vendor counters add a small, fixed number of tokens to each message. This overhead was measured and removed where it affects per-message comparisons. The 2023 GPT-4 tokenizer was also re-measured on the same data. It reproduces the 2023 study’s published figures, which serves as a check on the method. Prices are the vendors’ public input list prices on 19 October 2026 ^{8 9 10}.

The code, data and results are public in the companion repository ¹, so the numbers can be checked or re-run.

Bottom line

The rule that Vietnamese costs about twice as much as English fits the 2023 GPT-4 tokenizer (about 2.45 times) and, today, Claude (2.00 times). For the other vendors measured, the premium is about 1.24 to 1.40 times. Verified For a business choosing a model, the useful question is not “does Vietnamese cost more?” but “how much will my Vietnamese text cost on the models that answer well?” With the same list price, one model cost about 2.2 times as much as another at October 2026 prices — a gap that Google’s announced December price change would narrow to about 1.1 times.