Best AI for Turkish: four models tied in our 17 model test, and every proverb answer was wrong

Which AI is best for Turkish?

Four models tied for the top on Turkish, and none of them got more than 2 of 5. Our Multilingual LLM Bench gave 17 models the same 5 Turkish questions on 2 October 2026. Each question is worth 20 points on its own. Here's how 12 of the 17 did:

That's a low ceiling. Turkish averaged 23.5% across the 17 models, 28th of the bench's 30 languages. No model answered three of the five Turkish questions, and no other language had more than two that nobody solved.

You won't find a Turkish score on the big public boards 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 Turkish, as of October 2026.

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

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

Four models tied on 2 of 5, and one question decided the tie. 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 Turkish primary source, such as the Türk Dil Kurumu's dictionary or the Turkish Patent and Trademark Office's register. An answer in the wrong language scored zero.

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

Split the answers by question and the problem is plain. Fifteen of the 17 models knew one famous regional dish. Three questions beat all of them:

The Ottoman question is the one Turkish item we publish. It gives four terms in Arabic script, كاتب عدل among them, and asks for each one as the TDK's Güncel Türkçe Sözlük spells it today. The key is kâtibiadil, erkânıharp, müddeiumumi, vakayiname. Five models got all four. Seven of the 12 misses wrote the first term as two words or with a hyphen, such as kâtip adil or kâtib-i adil, where the dictionary prints one word.

That question is the whole tie. GPT Astra, Claude Opus, Gemini Pro and Gemini Flash got it and the dish. Claude Fable got the Ottoman spelling but missed the dish, so it sits on 1.

The two proverb questions describe an atasözü by its dictionary meaning and ask for the saying itself. All 17 models missed both. They also all missed the local name of a regional craft, a name the Turkish patent office records next to the official one.

Don't use our bench to rank Turkish against other languages. Turkish averaged 23.5% and German 64.7%, which says more about how hard each set of five questions was than about the language. Vietnamese went the other way, with three perfect scores: see the best AI for Vietnamese.

Is ChatGPT good at Turkish?

Yes, ChatGPT is good at everyday Turkish, but OpenAI doesn't publish a Turkish score for it. Its simple-evals results cover an MMLU test that professional translators rewrote in 14 languages, and Turkish isn't one of them.

Anthropic uses the same 14 languages on its multilingual support page, so Claude has no official Turkish 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 best independent number is older. In TR-MMLU, a test of 6,200 multiple choice questions that were drawn from the Turkish education system, GPT-4o scored 84.84%. Claude 3.5 Sonnet scored 84.40% and Gemini 1.5 Pro 76.74%, in a results paper dated August 2025. Every model we tested is newer than those three. In our own test, GPT Astra tied for first, while ChatGPT Sol and ChatGPT Luna got only 1 of 5 each.

Should you use a Turkish model like Kumru?

Use a Turkish model when you want to run it yourself, not as your everyday chatbot. The one independent test of them found the Turkish-only models behind the general ones. Cetvel, a Turkish benchmark posted on 22 August 2025, tested 33 open models of up to 70B parameters. Its authors found that "Turkish-centric instruction-tuned models generally underperform relative to multilingual or general-purpose models", and they named Llama 3 and Mistral as the stronger ones.

VNGRS says its Kumru is the exception. The model card says Kumru-2B was "developed for Turkish from scratch by VNGRS", trained on 500 GB of cleaned text for 300B tokens. It also says Kumru "overall surpasses significantly larger models" on Cetvel. Treat that as a lab's claim, since the scores sit in an image on the card.

You can download the 2B version under Apache 2.0, with an 8,192 token context. The larger Kumru-7B is a web demo at kumru.ai. Neither one appears in our bench, so we can't rank them against GPT or Claude.

What do ChatGPT, Claude and Gemini cost in Turkey?

ChatGPT Plus and Claude Pro both cost ₺999.99 a month in Turkey, and Google AI Pro costs ₺869.99. We took the ChatGPT and Claude prices from the Turkish App Store, where they're in-app prices, and the Google ones from Google's own Turkish plans page. We read all of them on 25 October 2026.

Skip the annual Claude plan. Anthropic lists it at ₺11,999.99 on the Turkish App Store, and that's 11 kuruş more than you'd pay for twelve months of ₺999.99 one at a time. Our page on Claude Pro prices by country compares it with other markets.

You don't need a card that pays abroad, either. On its payment methods page, Apple says that a Turkish Apple Account can pay by phone bill through Paycell, Türk Telekom Mobil Ödeme or Vodafone Mobil Ödeme, as well as with most credit and debit cards.

How should you prompt an AI in Turkish?

Write the prompt in Turkish, and say whether you want siz or sen. A prompt that names neither leaves the choice to the model, and the wrong one reads oddly in a work email or a note to a friend. Ask it to keep every Turkish letter (ç, ğ, ı, İ, ö, ş, ü). Then check each proverb against the TDK dictionary, because every model in our test missed both of ours.

```

Aşağıdaki koşullarla bir metin yaz.

Hitap: [siz / sen]

Uzunluk: yaklaşık [300] kelime, okur: [iş arkadaşları / öğrenciler]

Türkçe karakterleri eksiksiz kullan.

Bir atasözü kullanırsan anlamını parantez içinde yaz.

Emin olmadığın tarih, sayı ya da adları işaretle.

Konu: [konuyu buraya yaz]

```

For translation, paste the Turkish as it is and name the reader. This prompt protects the parts a model gets wrong:

```

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

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

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

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

[paste the Turkish text here]

```

Our guide to using AI to translate has more prompts and the checks to run after.

Where does Whizi fit for Turkish?

Whizi is our product, so here's where it helps and where it loses. Pro at $29.99 a month includes Claude Opus, Gemini Pro and Gemini Flash, three of the four models that tied first on our Turkish questions, plus ChatGPT Sol. Its side by side model chats run one Turkish prompt through two of them at once. GPT Astra, the fourth, needs Powerhouse at $49.99. Whizi loses on Turkish models: it doesn't carry Kumru, so for a model trained only on Turkish you'll need VNGRS's own download. Side by side model chats shows how compare works.