Best AI for Vietnamese: three of 17 models got all 5 of our Vietnamese questions right

Which AI is best for Vietnamese?

Three models got every Vietnamese question right, and no other language in our test had more than two do that. Our Multilingual LLM Bench gave 17 models the same 5 Vietnamese questions on 2 October 2026. Each question is worth 20 points on its own. Here's how 12 of the 17 did:

The gap after the top three is wide. Nobody else got more than 3, and three models got none. Across all 30 languages the bench recorded only nine perfect scores, and Vietnamese had three of them. German had two.

OpenAI, Anthropic and Artificial Analysis don't give you a Vietnamese score to check us against. OpenAI's and Anthropic's published multilingual tables cover 14 languages, and Artificial Analysis's Multilingual Index covers 16. None of them includes Vietnamese, as of October 2026. AI Singapore's SEA-HELM leaderboard does test Vietnamese, closed models among them, but it shows those scores only as charts.

For the wider fight between the three big assistants, see ChatGPT vs Claude vs Gemini.

How did 17 models do on Whizi's own Vietnamese questions?

GPT Astra, Gemini Pro and Claude Opus got 5 of 5, and five models tied behind them on 3. We gave every model the same 5 questions in each of 30 languages, at low reasoning. We graded each answer against a key from a Vietnamese source, such as a provincial newspaper or a Vietnamese dictionary. An answer in the wrong language scored zero.

Five questions is a small sample. One question moves a model's Vietnamese score by 20 points, so read this as a rough order, not a final ranking.

Split the answers by question and you can see where the top three pulled away. Most models handled the calendar and the local facts. The proverbs split the field:

The festival question is the one Vietnamese item we publish. It asks for the other name that people give to the Trò Trám festival, which is held on the 11th and 12th of the first lunar month in Phùng Nguyên, Phú Thọ. Báo Phú Thọ, the provincial paper, gives it as Linh tinh tình phộc. Ten of the 17 models wrote it. Gemini Flash wrote the right name inside quotation marks, and our strict grader marked it as wrong.

The two proverb questions describe a thành ngữ by its dictionary meaning and ask for the four word saying itself. GPT Astra, Gemini Pro and Claude Opus got both. Grok, Qwen Max and ChatGPT Luna got one each, and the other 11 models got neither.

That's still far better than most languages managed. Our 17 models got 9 of 34 Vietnamese proverb answers right, against 0 of 34 in Turkish (see the best AI for Turkish) and 1 of 34 in Korean. Don't read that as Vietnamese being easy, though. It says as much about how hard each set of five questions was as about the language.

Is ChatGPT good at Vietnamese?

Yes, ChatGPT is good at Vietnamese, and its top model is one of the three that got all 5 of our questions. GPT Astra scored 5 of 5, while ChatGPT Sol and ChatGPT Luna got 3 each. OpenAI doesn't publish a Vietnamese score itself. Its simple-evals results cover an MMLU test that professional translators rewrote in 14 languages, and Vietnamese isn't one of them.

Anthropic uses the same 14 languages on its multilingual support page, so Claude has no official Vietnamese number either. Its advice for every other language is short: "Claude is capable in many languages beyond those benchmarked in the following table. Test with any languages relevant to your specific use cases."

The main Vietnamese exam benchmark hasn't scored an OpenAI model since 2024. In the zero-shot results of VMLU, GPT-4 scored 65.53, which was just behind the 66.44 of Llama-3-70B. The newest OpenAI row on its leaderboard is GPT-4o mini at 62.87, and it was evaluated in August 2024. There's no Claude model on it at all.

Which AI is best for Vietnamese translation?

For Vietnamese translation, start with GPT Astra, Claude Opus or Gemini Pro, the three models that read our Vietnamese questions best. We didn't find a recent human graded test of English to Vietnamese for these chatbots, so our test is the closest check. Paste the Vietnamese with its diacritics. Anthropic's guide says: "Submit text in its native script rather than transliteration for optimal results." Then use a prompt that protects the parts machines get wrong:

```

Translate this Vietnamese text into natural English for [a business email / a school report / subtitles].

Keep every name, number, date and amount exactly as written.

If a Vietnamese proverb or idiom has no English match, translate its meaning and put the Vietnamese in brackets.

At the end, list any word you were unsure of.

[paste the Vietnamese text here]

```

For English to Vietnamese, swap the direction and say who the reader is, since that sets the pronouns (more on that below). Our guide to using AI to translate has more prompts and the checks to run after.

Should you use a Vietnamese model like VAI-LLM or V-LLM?

Use a Vietnamese model when you're building on one, not as your everyday chatbot, because no public test puts them next to a current GPT, Claude or Gemini model. VMLU is where they compete. ZaloAI and JAIST built it, and the VMLU site calls it "a human-centric benchmark suite specifically designed to assess the overall capabilities of foundation models". Its repository counts 10,880 multiple choice questions across 58 subjects.

Here's the best model from each of the top six labs on its fine-tuned model leaderboard, read on 25 October 2026. Older versions from VAI and VinSmart Future fill the ranks in between:

Those scores have climbed fast. In VMLU's original table, the best fine-tuned model, VNPTAI.IO-14B, scored 55.83. The board doesn't say which of the new leaders are open or how you'd use them, so check each lab's own page before you plan on one.

What do ChatGPT, Claude and Gemini cost in Vietnam?

ChatGPT Plus costs ₫499,000 a month in Vietnam, Claude Pro costs ₫599,000, and Google AI Pro is ₫489,000. We took the ChatGPT and Claude prices from the Vietnamese App Store, where they're in-app prices, and the Google ones from Google's own Vietnamese plans page. We read all of them on 25 October 2026.

Claude Pro is the dearest of the three main plans. Anthropic charges ₫100,000 a month more for it than OpenAI does for ChatGPT Plus. Its annual plan at ₫6,999,000 saves ₫189,000 on twelve monthly payments, about 2.6%.

You don't need a card that pays abroad. On its payment methods page, Apple says that a Vietnamese Apple Account can pay with MoMo, ShopeePay, VNPAY or Zalopay, as well as with most credit and debit cards. Our page on Claude Pro prices by country compares other markets.

How should you prompt an AI in Vietnamese?

Write the prompt in Vietnamese with full diacritics, and say how the text should address the reader. Vietnamese has no neutral "you": the model has to pick between bạn, anh chị, quý khách and others. If you don't name one, it picks for you. Say whether you want northern or southern wording too, since a reader in Hà Nội and one in Sài Gòn notice it.

```

Hãy viết một bài theo các yêu cầu sau.

Xưng hô: [quý khách / bạn]

Cách dùng từ: [miền Bắc / miền Nam]

Độ dài: khoảng [300] chữ, cho [đồng nghiệp / học sinh]

Viết tiếng Việt có dấu đầy đủ.

Nếu dùng thành ngữ, hãy ghi nghĩa trong ngoặc.

Đánh dấu ngày tháng, con số hoặc tên mà bạn không chắc.

Chủ đề: [ghi chủ đề ở đây]

```

Then check every proverb, place name and date by hand. Most of the 17 models in our test missed at least one of two Vietnamese sayings, and their wrong answers looked just as sure as the right ones.

Where does Whizi fit for Vietnamese?

Whizi is our product, so here's where it helps and where it loses. Pro at $29.99 a month includes Claude Opus and Gemini Pro, two of the three models that got all 5 of our Vietnamese questions, plus ChatGPT Sol, which got 3. Its side by side model chats run one Vietnamese prompt through two of them at once. GPT Astra, the third perfect scorer, needs Powerhouse at $49.99. Whizi loses on Vietnamese models: it carries none of the VMLU leaders, so for VAI-LLM or V-LLM you'll need the labs themselves. Side by side model chats shows how compare works.