Best AI for Korean: GPT Astra won our 17 model test, and 33 of 34 proverb answers were wrong

Which AI is best for Korean?

Our Multilingual LLM Bench gave 17 models the same 5 Korean questions on 2 October 2026, and four of them tied on 3 behind GPT-6 Astra. There are only 5 questions, so each one is worth 20 points on its own. Here's how 12 of the 17 did:

The public board points the same way, with one more name in it. Artificial Analysis runs a multiple choice test in Korean, and on 18 October 2026 its top five were Claude Opus 4.6 (max) at 93, then Gemini 3.1 Pro Preview and three more Claude Opus versions at 92. No GPT model was in its top five, and neither was any GPT-6 or Claude Opus 5 model we tested.

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

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

GPT-6 Astra led alone with 4 of 5, and four models tied behind it 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 that we took from a Korean primary source, such as the dictionary of the National Institute of Korean Language or a local ordinance. An answer in the wrong language scored zero.

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

Split the answers by question and the pattern is plain. Local facts were fine. Proverbs weren't:

The two proverb questions give a 속담 by its dictionary meaning and ask for the saying itself, in the exact form the Standard Korean Language Dictionary prints. The first one is public. It describes pestering someone for an unrelated favour while everyone else is frantic. The answer is 상여 나갈 때 귀청 내 달란다. No model wrote it. Six models guessed a saying that begins 불난 집에, a house on fire. None of those matches the dictionary entry.

GPT-6 Astra was the only model to get the second proverb, and that one answer is its whole lead. The local questions caught a different mistake. Five of the 15 wrong answers there were the date from an older ordinance that a city or province had since replaced. The models knew a fact. They knew the old one.

Don't use our bench to rank Korean against other languages. Korean averaged 30.6% across the 17 models, 23rd of 30, and Japanese averaged 40.0%. That says more about how hard each set of five questions was than about the language.

Is ChatGPT good at Korean?

Yes, ChatGPT is good at standard Korean, and it's the AI most Koreans use. OpenAI's simple-evals results put its o3 reasoning model at 0.893 in Korean, on an MMLU test that professional translators rewrote in 14 languages. GPT-4.1 mini scored 0.801, so the smaller ones fall behind it. In our test, GPT-6 Astra led and GPT-6 Sol got 3 of 5, while GPT-6 Luna, the cheapest GPT we ran, got 1.

Korea's Ministry of Science and ICT counts which of them people use. Its 2025 Internet Usage Survey, published on 31 March 2026, found that 44.5% of Koreans had used a generative AI service, up from 33.3% a year before. The ministry's release names the main services as "챗지피티(41.8%), 제미나이(9.8%), 코파일럿(2.2%), 클로바 X(2.0%)". Kakao says ChatGPT for Kakao, its version inside KakaoTalk, had passed 11 million cumulative subscribers by May 2026.

Anthropic publishes the same translated test for Claude. Its multilingual support page says Claude Sonnet 4.5 kept 96.7% of its English score in Korean, and Claude Haiku 4.5 kept 93.3%.

Which AI is best for Korean translation?

Gemini 2.5 Pro was the best major AI for English to Korean translation in WMT 2025, the main recent test graded by people. The newer Gemini 3.1 Pro wasn't in it. In the WMT 2025 general translation task, professional annotators marked the errors in each system's English to Korean output. The scores are error penalties, so closer to zero is better. The human reference scored -1.9, a 7B research system built on Hunyuan -2.5, and Gemini 2.5 Pro -2.7. All three sat in the top cluster, ranked 1 to 3. GPT-4.1 (-3.3) and Claude 4 (-3.4) came next, and Qwen3 235B scored -5.5.

Papago and DeepL weren't in that evaluation, so the study can't rank them. When you use a chatbot, paste the Korean in Hangul, and tell it the speech level you want. 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 Korean text into natural English for [a business email / a school report / subtitles].

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

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

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

[paste the Korean text here]

```

For English to Korean, swap the direction and name the speech level: 합쇼체 for a formal letter, 해요체 for most emails, 반말 only for friends. Our guide to using AI to translate has more prompts and the checks to run after.

What do ChatGPT, Claude and Gemini cost in Korea?

ChatGPT Plus costs ₩29,000 a month in Korea, Claude Pro ₩33,000 and Google AI Pro ₩29,000. Each row below is the in-app price on the Korean App Store, read on 18 October 2026.

You don't need a card that pays abroad. Apple's payment methods page lists Kakao Pay, Naver Pay, PAYCO, Toss Pay, phone billing on KT, LG U+ and SK Telecom, and most credit and debit cards for a Korean Apple Account. On the web, OpenAI lists KRW among its billing currencies. Our page on ChatGPT payment methods covers what its own checkout takes in Korea.

Claude Pro is the dearest of the three main plans. Anthropic charges ₩33,000 for it, which is ₩4,000 more than you'd pay for ChatGPT Plus. If you buy its annual plan at ₩349,000, that works out at about 12% less than twelve monthly payments.

Should you use a Korean model like EXAONE, Solar, A.X, HyperCLOVA X or Kanana?

Use a Korean model when you want to run it yourself or build on it, not as your everyday chatbot. Korea's own labs now publish open models that you can download, but the scores on them are mostly the labs' own, and two of the big Korean chat apps have closed. Naver shut CLOVA X and its search assistant Cue: on 9 April 2026. Kakao ended its standalone Kanana app on 15 October 2026 and folded it into KakaoTalk.

The government set out to narrow them to two. In its sovereign AI foundation model project, each team is scored out of 100: 40 points come from the benchmarks, 35 from the experts and 25 from the users. In the first round, on 15 January 2026, LG AI Research scored highest at 90.2, and the ministry cut Naver Cloud and NC AI. Its briefing said Naver's model didn't meet the rule that a model must be built independently. Motif Technologies, which wasn't in the first round, joined it for the second. On 19 August 2026 the ministry put SK Telecom, Upstage and LG AI Research through, and it cut Motif. It didn't publish the scores, and it said that the plan to end with two teams is now under review. Newspim, a Korean news agency, reported that SK Telecom had 70.6, Upstage 69.9 and LG AI Research 69.0.

Each lab grades itself, so read their claims as claims. LG says K-EXAONE 2.0 averaged 70.1 across 24 benchmarks, up from 63.3 for its first version. Upstage's model card gives Solar Open 2 a 78.4 on KMMLU-Pro, and marks three of its Korean scores as in-house tests. SK Telecom's A. (에이닷) app is the one Korean assistant with scale: SK Telecom said it passed 10 million monthly users in September 2025, counting 5.5 million who use it inside calls, B tv and TMAP.

Which Korean AI benchmarks can you trust?

KMMLU is the benchmark to trust first, because it's built from Korean exams and not translated from English, and Upstage still reports its newer KMMLU-Pro version. Each of the tests below measures something different, so a model can lead one of them and trail in another. Gemini 3.1 Pro is the plain example: second on Artificial Analysis, 2 of 5 on ours.

The KMMLU paper ended on a line that still holds for the hardest Korean questions. Its authors found "the best public model to score 50.5%, leaving significant room for improvement." Frontier models now score far higher on multiple choice, up to 93 on Artificial Analysis's Korean index. They still couldn't write a dictionary proverb from its meaning.

How should you prompt an AI in Korean?

Write the prompt in Korean, and say which speech level you want. A prompt that names no level leaves the choice to the model, and a polite 해요체 reads oddly in a formal report or a chat with friends. Ask for Hangul only, unless you want Hanja next to a word. Then check every proverb, place name and date by hand, because our test caught models giving a replaced date with full confidence.

```

다음 조건으로 글을 써 주세요.

문체: [합쇼체 / 해요체 / 반말]

분량: 약 [500]자, 독자: [회사 동료 / 고등학생 / 일반 독자]

속담이나 관용구를 쓰면 괄호 안에 뜻을 적어 주세요.

확실하지 않은 날짜나 수치는 따로 표시해 주세요.

주제: [주제를 여기에]

```

The same prompt works for a cover letter (자기소개서) or a blog post: swap the topic line and the reader. For school work, ask for the reasoning, not a finished answer to copy.

Where does Whizi fit for Korean?

Whizi is our product, so here's where it helps and where it loses. Pro at $29.99 a month includes Claude Opus, ChatGPT Sol and Gemini 3.8 Flash, three of the five models that got 3 or more of our Korean questions, plus Gemini 3.1 Pro. Its side by side model chats run one Korean prompt through two of them at once. GPT-6 Astra, the model that led, needs Powerhouse at $49.99. Starter at $15.99 has ChatGPT Luna, which got 1 of 5, and Claude Haiku. Whizi loses on Korean models. Upstage's Solar Pro 4 is the only one in its catalog, so for EXAONE, A.X or HyperCLOVA X you'll need the labs' own downloads. Side by side model chats shows how compare works.