Best AI for French: 17 models tested, GPT Astra won with 4 of 5, and none knew our Belgian idiom

Which AI is best for French?

Four of the five French questions in our Multilingual LLM Bench, run on 2 October 2026, are about Belgium or Switzerland, not France. Every one of the 17 models got the same five. Only two knew the Swiss idiom: GPT Astra and Claude Fable. Here's how many models got each one right:

The one public French question asks for the date of a Geneva holiday, the anniversary of the restoration of the Republic. Geneva's law on public holidays lists it as "31 Décembre, anniversaire de la restauration de la République". All 17 of the models answered 31 décembre. You can read the question and every answer on our public question page.

The other four stay private, so we can rerun them on new models. The Swiss idiom decided the top. GPT Astra got it, plus the grammar question and both local facts, the same three that eight other models got. Claude Fable also knew the Swiss idiom, but it missed the grammar question.

This page ranks the chatbots that write, translate and answer in French. If you want to learn French, that's a different pick: tutor apps such as Langua and Talkpal are built around speaking practice, and none of them was in our test.

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

GPT Astra won with 4 of 5, and nine models shared second place with 3. 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 French score by 20 points, so read this as a rough order, not a fine ranking.

Claude Sonnet came last in French with 1 of 5. It missed the grammar question that 15 of the 17 got right. Grok went the other way. It got none of our Spanish questions, but 3 of 5 in French. GPT Luna, the cheapest model we ran at $0.01 for all 150 questions, got 2.

Don't read the France gap as a verdict on France French. None of our five questions asks about France, so this test says nothing about how well a model writes Parisian French. We didn't test that. It says how often a model knows the French of Belgium and Switzerland.

Don't use our bench to rank French against other languages either. French averaged 51.8% across the 17 models, against 25.9% for Spanish and 35.3% for Portuguese. That says more about how hard each set of five questions was than about the language.

What do the public French leaderboards say?

The public leaderboards put all of the big models close to the top in French, and they can't split the leaders. Artificial Analysis scores French on its Multilingual Index. As we read it on 25 October 2026, Gemini 3.1 Pro Preview and GPT-5.1 (high) tie for first at 94, with Gemini 3 Pro Preview, GPT-5 (high) and Claude Opus 4.6 (max) at 93. A gap of one point is noise. The board is also behind the releases: Claude Opus 5.5, the GPT-6 models and Gemini 4 weren't on it when we looked.

The vendors publish their own numbers too. Anthropic's multilingual support page says Claude Sonnet 4.5 keeps 97.5% of its English score in French, and that Claude Haiku 4.5 keeps 95.7%. French is fourth of the 14 languages in that table, behind Spanish, Italian and Portuguese. Anthropic says that the tests were "translated into 14 additional languages by professional human translators".

OpenAI publishes its own numbers on the same translated MMLU test, in its simple-evals results. There, its o3 model scores 0.906 in French. GPT-4.1 nano, the lowest scorer on that page, got 0.739.

All of these are translated tests. A translated question can't ask for a Belgian idiom, and that's the gap our five questions aim at.

Does AI write France, Quebec, Belgian or Swiss French?

AI writes the French that your prompt asks for, so name the country. The words and the numbers are where your reader will notice it first. In Quebec, the language office, the OQLF, lists fin de semaine as its term for the weekend, and it says "dès les années 1920, c'est le terme fin de semaine qui s'est imposé dans l'usage". In France, week-end is the normal word, and the OQLF itself notes that European French accepts it.

Numbers are the second split. The Trésor de la langue française marks septante, for seventy, as a word of Belgium, of French speaking Switzerland and of the eastern edge of France. In Paris, it's soixante-dix. A model that writes soixante-dix in a Brussels invoice isn't wrong, but your reader will hear France in it.

Register is the third. Pick tu or vous before you start, and tell the model, because a draft that drifts between the two reads as careless in every French speaking country.

```

Write in French for readers in [France / Quebec / Belgium / Switzerland].

Use [tu / vous] for the reader, and keep it the same all the way through.

Use the words people in that country use, for example [fin de semaine or week-end] and [septante or soixante-dix].

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

At the end, list any word or phrase that would sound foreign in that country.

Task: [a product email / a school essay / a LinkedIn post] about [topic], about [200] words.

```

For translation, swap the task line for "Translate this English text into French for [country]" and keep the rest. Our guide to using AI to translate covers the checks to run after.

Is Mistral or a French open model better at French?

We can't say Mistral is better at French, because no test we found shows it. Mistral's models weren't in our run, and none of them is in Artificial Analysis's French top five as we read it on 25 October 2026. Mistral sells its Pro plan at $14.99 a month before tax, with a $5.99 student rate, per its pricing page. That's less than ChatGPT Plus, Claude Pro or Google AI Pro in every store below. If you're curious, give Mistral's chat and one of the models above the same French prompt and compare the drafts.

Lucie-7B is the open model built around French. OpenLLM-France released it with a technical paper on 15 March 2025. Its model card lists 6.7 billion parameters, an Apache 2.0 licence and 3 trillion training tokens, of which French is 32.4% and English 33.2%. It's a model you run yourself. Neither the paper nor the card compares it with Claude Opus, GPT or Gemini, and it wasn't in our run.

How much do ChatGPT, Claude and Gemini cost in France, Belgium, Switzerland and Canada?

Canada is the cheapest of the four for ChatGPT Plus: it's C$24.99 a month there, which is $17.84. We read each app's App Store listing in France, Belgium, Switzerland and Canada on 25 October 2026. The dollar figures use the Currency API rates of 24 October 2026, rounded to the cent.

France and Belgium charge the same euro price for all six plans. Google AI Pro is the cheapest of the three main plans in France, Belgium and Switzerland. In Canada, ChatGPT Plus is. Canada's Google AI Plus is a 200 GB plan, while the European stores sell a 400 GB one.

If you want GPT Astra, the model that topped our test, you won't get it in ChatGPT Plus chats. OpenAI's help page says "GPT-6 Pro, powered by GPT-6 Astra, is available in ChatGPT for Pro $100, Pro $200, Business, and Enterprise plans." Plus gets GPT Sol in chat, and it gets Astra only in ChatGPT Work and Codex. Claude Pro does include Claude Opus, as claude.com/pricing shows. For more countries, see Claude Pro price by country. If your card keeps failing, paying for AI when your card is declined covers what you can do about it.

Where does Whizi fit for French?

Whizi is our product, so here's where it helps and where it loses. GPT Astra, the only model with 4 of 5 in our French test, is on Whizi Powerhouse at $49.99 a month, less than half the price of ChatGPT Pro 100 in all four stores above. Credits per message vary by model, and Astra is one of the dearest, so Powerhouse suits checking hard French, not chatting all day. Pro at $29.99 has Claude Opus, ChatGPT Sol and Gemini Pro, three models that each got 3 of 5, with side by side model chats to compare them on one prompt. Whizi loses on a single plan: Claude Pro alone costs less in all four stores, and Lucie isn't in Whizi's catalogue. Side by side model chats shows how compare works.