Best AI for Japanese: Claude Opus overall, Gemini for translation, and only 4 of 17 models knew Gunma's 2024 karuta card

Which AI is best for Japanese?

Claude Opus tops two of the four Japanese tests we could find, one of them in a tie, and Gemini tops the other two. The sharpest result is a local one: only 4 of the 17 models we ran knew the current wording of a card in Gunma's karuta game, which changed in 2024. We read all four tests on 18 October 2026, and each row below is that test's own result.

Nejumi is the Japanese leaderboard run by Weights & Biases Japan. We read its scores through Qualiteg's summary of the 1 September 2026 snapshot. Weights & Biases Japan launched version 4 on 27 August 2025 and said it aims to give companies "より実践的な指標" (more practical measures) for choosing a model. The four boards test different model generations and different tasks, so it's no surprise that they disagree at the top. Our own run is the only one of them that tested GPT-6 and Claude Opus 5.5:

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

GPT Sol and Claude Opus tied first, with 4 of 5 Japanese questions right. That's from our Multilingual LLM Bench, run on 2 October 2026. Every model got the same 5 questions in each of 30 languages, at low reasoning. We graded each answer against a sourced key, and an answer in the wrong language scored zero.

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

The five questions split three ways. A question on the traditional Japanese calendar was easy: 12 of 17 models got it. One local fact was easy too, at 13 of 17. The two sayings, each described only by its meaning, were the wall. Claude Opus was the only model to name the first one, and 4 models named the second.

The public question is the one we'd remember. It asks for the exact wording of the "chi" card in Jōmō Karuta, Gunma Prefecture's local card game, on sets shipped from April 2024. Gunma Prefecture changed that card in January 2024, after its population fell to 1,899,297 on 1 December 2023. Here's what each model wrote:

So 4 of 17 knew the new card, and 4 more knew the old one. The old wording ran for 31 years, so it fills the training data. A model that answers a Japanese local question with confidence may be a few years behind. Check dates, prices and anything local by hand.

Don't use our bench to rank Japanese against other languages. The 17 models averaged 40.0% in Japanese and 61.2% in English, but that says more about how hard each set of five questions was than about the language.

Is ChatGPT good at Japanese?

Yes, ChatGPT is good at Japanese. OpenAI's simple-evals results put its o3 reasoning model at 0.890 in Japanese, on an MMLU test that professional translators rewrote in 14 languages. The small models fall back fast: GPT-4.1 mini scored 0.810 and GPT-4.1 nano 0.690. In our test, GPT Sol tied first and GPT Luna, the cheapest GPT we ran, got 2 of 5.

Anthropic publishes the same test for Claude. Its multilingual support page says Claude Sonnet 4.5 kept 96.8% of its English score in Japanese, and the smaller Claude Haiku 4.5 kept 93.5%. That's close to its Korean and Chinese numbers. For the big chatbots, Japanese is a first class language.

Which AI is best for Japanese translation?

Gemini Pro is the best AI for English to Japanese translation on the one recent test graded by people. The WMT 2025 findings had professional annotators score every system from 0 to 100. Here's the English to Japanese table:

Two limits apply. WMT 2025's only pair out of Japanese went into Chinese, so it has no human graded result for Japanese to English. And the models are a year old: no GPT-6 or Claude Opus 5 model was in it. The organisers also found Gemini 2.5 Pro in the top group for 14 of 16 language pairs, so Gemini's lead isn't a Japanese quirk.

Paste kana and kanji, not romaji, when you can. Anthropic's guide says it plainly: "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 Japanese text into natural English for [a business email / a manga page / a contract summary].

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

Tell me the politeness level of the original (casual, です/ます, or keigo).

If a phrase has no English match, translate its meaning and put the Japanese in brackets.

[paste the Japanese text here]

```

For English to Japanese, name the register: casual, です/ます, or full keigo for a client. If you want a dedicated Japanese tool, Preferred Networks runs PLaMo翻訳 on its own PLaMo models, with a free plan of 50,000 characters of text a month. Our guide to using AI to translate has the checks to run after.

Which AI is best for learning Japanese?

Claude Opus or GPT is the best AI for learning Japanese, because both tied first on our questions and both can explain the grammar in English. The catch is the sayings. Only Claude Opus named both of our Japanese sayings, so check any kotowaza a tutor bot hands you against a dictionary.

Tell the model your level and ask it to correct you, not to talk at you:

```

You are my Japanese tutor. I am around JLPT N4.

Talk to me in simple Japanese with furigana in brackets after each kanji.

After each of my replies, correct my mistakes in English, then continue the conversation.

Use です/ます unless I ask for casual speech.

Topic: ordering food at a ramen shop.

```

For speaking practice, use a chatbot's voice mode. Google lists Japanese among the languages of the Gemini mobile app, with no limits marked against it.

Should you use a Japanese made AI like Sakana, PLaMo or Sarashina?

Use a Japanese made AI when your data has to stay in Japan or on your own servers, not for raw quality. No Japanese built model is near the top of Nejumi yet. In Qualiteg's read of the 1 September 2026 snapshot, the best one, GPT-OSS-Swallow-120B from the Swallow team at Institute of Science Tokyo, scored 0.6914. The 50th model on the board scored 0.7954.

Sakana AI makes the strongest case for a Japanese tune. Its Namazu models take big open models and train them for Japan. Sakana says the base DeepSeek V3.1 Terminus refused 72% of its Japan related questions on politics, history and diplomacy, and its Namazu version brought that close to zero. That's Sakana's own test.

Each lab grades itself, so read the scores as claims. ELYZA's own model card is honest about it. Its new 33B model averages 61.69 on the card's own benchmark set, and the same card puts Google's Gemma 4 31B ahead at 70.54. The rinna team reports 8.53 on Japanese MT-Bench for its Bakeneko model. But MT-Bench is graded by another AI, and the big chatbots don't publish a score on it. NTT says tsuzumi 2 is "世界トップクラス" (world class) against models of its size, without a public number.

How much do ChatGPT, Claude and Gemini cost in yen?

ChatGPT Plus and Claude Pro both cost ¥3,000 a month in the Japanese App Store, and Google AI Pro costs ¥2,900. We read these prices on 18 October 2026:

Google's own Japanese plans page shows ¥725 for AI Plus, ¥2,900 for AI Pro and Ultra from ¥14,500 a month. On the website, Anthropic has added 10% Japanese consumption tax to every Claude plan since 1 April 2026. Our Claude Pro prices by country page compares the app and web prices.

The app stores are also where Japanese wallets work. Apple's payment methods page lists cards, Apple Pay, PayPay, Merpay, au, NTT DOCOMO and SoftBank phone billing, and Apple Account balance for Japan. Google Play adds JCB, Edy, PayPal and Rakuten billing. It also warns: "You can't use cash to pay for subscriptions." So convenience store cash only works once it's loaded as Play balance or a gift card. claude.ai itself takes cards only, as our Claude payment methods guide shows.

Where does Whizi fit for Japanese?

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 Pro, the three families that top the Japanese tests above. Its side by side model chats run one Japanese prompt through two of them at once. Buyers in Japan are charged in yen at checkout, and the interface has a Japanese setting. Sakana Namazu is in the catalog too, on Powerhouse at $49.99. Whizi loses on the rest of the Japanese field: it carries no PLaMo, Sarashina or tsuzumi model, and if you live in Japan, Sakana Chat is free.