Best AI for translation
Which AI translates best? Nobody has published per-language proof worth trusting. Six steps that let you check a translation yourself instead.
You paste a paragraph, name a language, and something fluent comes back. A fluent result still needs checking. Fluency is the one quality you can judge without knowing the target language, and the one that tells you least about whether your meaning arrived. A sentence that flips a negation reads exactly like one that does not.
So the question worth answering is how you catch the error before your reader does.
Why there is no per-language ranking in this article
We carry 59 models from 17 providers. We have not evaluated them language by language, and we are not going to imply an order we cannot show you.
"Model X is best at Japanese" is an easy sentence to write and a hard one to earn. It needs a test set, native-speaker raters, a defined text type and a published method. Look for that measurement when a listicle makes the claim. It is usually missing.
What is checkable: models are trained on different mixtures of text by labs in different countries, and they do behave differently on the same passage. You can see that in ten minutes with your own text, and your language pair and subject are the only test conditions you care about. What a chat model adds over a translation box is instructions and follow-up questions, which is the method below.
This is not a niche audience. In our own usage data for the 90 days to 2 August 2026, English accounted for roughly half the messages we could attribute to a language, and Spanish was the largest non-English share by a wide margin.
Nothing routes you to a model by language. No part of the app reads your text to decide which model should answer. The one place language detection runs at all is read aloud, which sniffs the text so a French reply is not spoken with an English voice. Model choice is yours on web and on mobile, which is why the steps below are worth running.
Step 1. Tell it what the text is for
Most bad AI translations get the words right and the register wrong, because the model was handed a paragraph with no idea what job it does. Before your text, write four short lines:
Purpose. A refund request, a birthday message, a notice to tenants.
Reader. A support team, your partner's parents, an official you have never met.
Register. Formal or warm, and if the language distinguishes them, which second person you want: tú or usted in Spanish, tu or vous in French.
Destination. An email, a WhatsApp message, a printed sign, a letter that has to sound official.
Then paste the text. The reader line does the most work, because it constrains the others.
Step 2. Ask for a literal back-translation in the same reply
Add one sentence to your request: "Then translate your own translation back into English, literally, without smoothing it."
Read that as though you had never seen your original. Look for four things: a clause that disappeared, a number or date that changed, a negation that flipped, a name translated when it should have stayed put. All four are visible in your own language.
A back-translation from the same model has a blind spot. It knows what it meant, so it can quietly restore the sense it intended instead of the sense it wrote. It catches dropped and inverted meaning well. It will not tell you whether the result sounds like a person, which is what Step 4 is for.
Step 3. Ask what it was unsure about
This question produces the most useful paragraph of the exchange: "List anything you were unsure about, with the alternatives you rejected and why."
A good answer names specifics: the word for "manager" in this country and this industry, whether your greeting is too familiar for a first email, an idiom rendered as its meaning, a brand name that reads as a common noun. You may not know which option is right, but you will know which ones matter, and now you have three words to ask a colleague about instead of a whole page.
Step 4. Take the same passage to a model from a different maker
Two models from one lab share training decisions and often share mistakes. Two from different makers, say a Gemini model and a Mistral one, or Claude and Qwen, are far more likely to disagree exactly where your text is ambiguous. Disagreement is the signal, and where both land on the same wording you can relax about that sentence.
Changing model inside one conversation earns its keep here, because the second model needs your whole set-up and retyping it is how people skip the check. Use the branch action on the message: it copies the conversation up to that point into a new thread and re-answers with the model you pick, leaving the original as it was.
The labels are swapped between our platforms, so read this before you tap. On web, "Try Again" is the branching action, the safe one, and "Re-generate" replaces the existing answer. On mobile the words swap over: "Branch here" is safe and "Try again" replaces the answer in place. Take the branch on either platform to keep both translations.
The branch opens as its own conversation, so you compare by moving between two threads. No view puts two answers next to each other.
Step 5. Lock the words that must not move
Once you have seen where the two models disagree, write a five-line glossary and re-run the translation with it attached: names of people and companies, product names, job titles, anything you used in earlier correspondence with this reader. Say which items stay in the source language: models translate brand names surprisingly often, and a customer who only knows the English name will not recognise it translated.
Step 6. Decide what still needs a person
Some of this should not be happening in a chat window at all.
Send it to a human translator if it is legal, medical or contractual. Contracts, immigration paperwork, medical instructions, safety notices, anything sworn or certified. An error there lands on your case file or someone's dosage, and a certified translator carries professional responsibility no model does.
For everything else, these steps get you a translation you can defend, and two minutes from a native speaker gets you one you can trust.
What a file loses on the way through
Attach a file instead of pasting a paragraph and the limits of a chat window arrive quickly.
Nearly every model in the app can read a plain text file, though attaching one needs a paid plan. PDFs are narrower: 24 of the 59 models accept a PDF upload, and the picker blocks the attach on the rest. A photo of a menu or a sign goes to models that take image input, and many will read the text in it, though that is the provider's doing and not something we would tell you to rely on.
Layout survives none of those routes. You get translated text back, and tables, footnotes and column breaks stay behind, so a letter pastes into your own template and a formatted brochure means rebuilding it.
Worth knowing before you paste correspondence that is not yours: we do not train on your conversations, and what the provider behind a given model does varies. Six of the 59 carry a notice on the info card. Grok's states outright that "xAI may use your conversations to train Grok models". The other 54 are silent.
Try this tonight
Take one paragraph you already know the answer to, something professionally translated or checked by a friend, and run it through Steps 1 to 4 on two models from different makers. Twenty-two of our 59 models are open on the free plan, enough to run that comparison before you decide anything about a subscription, and the plans are laid out in our pricing section.
You will learn more about your own language pair in that quarter of an hour than any ranked list can tell you.