Best AI for Translation: What the Only Independent Test Found

WMT25 is the annual machine translation evaluation whose results are graded by human annotators rather than by software. In it, human translators landed in the winning cluster for only six of the 15 language pairs where their work was assessed. The best machine system in that same evaluation placed in the top cluster for 14 of 16 pairs, and matched or surpassed the human translation on 10 of them. That is the strongest independent evidence anyone has on which is the best AI for translation, and almost none of the pages ranking for the question mention it exists.

What follows answers which model to open for your language pair and your kind of text, based on that evaluation rather than on a vendor scoreboard. It also explains why the benchmarks you find on Google contradict each other so sharply, and what Apple's translation actually does on an iPhone and on a Mac. Several widely repeated claims in this space collapse the moment you follow them back to their source, and those are flagged as we go.

The Key Takeaways

• Gemini won the only hand-graded test: at WMT25 it was the best system overall, in the top cluster for 14 of 16 language pairs, ahead of GPT and Claude.
• Human translation is no longer the ceiling: professional human references reached the winning cluster in only six of 15 pairs, and the top machine system matched or beat them on 10 of 16.
• The evidence is a year old: that evaluation tested Gemini 2.5 Pro, GPT-4.1 and Claude 4. All three have since been replaced, so treat the result as a verdict on model families, not on today's version numbers.
• Almost every ranking you will find is written by a seller: the pages competing for this query are published by translation companies that place their own product first.
• Apple's version is the private one: Live Translation runs entirely on your iPhone, so nothing is sent to a server, but it covers nine languages and needs an iPhone 15 Pro or later.

The Best AI for Translation, by Job

There is no single winner, and any page that gives you one without qualification is selling something. What the evidence supports is a split by task. Here is the short version, with the reasoning underneath.

If you want the current specifications behind those names, we keep them updated on the best AI models rankings page, and individually for Gemini 3.7 Flash, Claude Opus 5 and GPT-5.6.

What the Only Independent Benchmark Found

Most translation comparisons you will read are one person pasting a paragraph into four tools and judging the output themselves. The Conference on Machine Translation runs the exception. Its general translation task puts every system in front of human annotators who mark the actual error spans in each output, a protocol called Error Span Annotation, across dozens of language pairs and several content domains. The organisers describe that pool as professional annotators, while also noting they were not specialists in this style of annotation.

The WMT25 findings paper evaluated 60 systems in total, 36 submitted by research teams and 24 collected by the organisers from commercial models and online translation services. Its conclusion is stated plainly: the best system overall was Gemini 2.5 Pro, which placed in the top cluster for 14 language pairs.

The ranking

The paper publishes no tidy leaderboard of consumer models, but its average human ranks across the evaluated pairs put Gemini 2.5 Pro clearly first on 2.28, then GPT-4.1 on 4.40, then Claude 4 on 6.62 and DeepSeek V3 on 7.14. Behind them sat a cluster including Mistral Medium, Llama 4 Maverick, Command A, Qwen3 and both Gemma models, none of them close. The gaps between the top four were meaningful but not enormous. The gap to the bottom of the field was. The best system on a given pair usually scored 90 or above, while mid-tier models often scored 60 or below, and the weakest could score zero.

The finding nobody quotes

The result that should reframe this whole question concerns the humans. Professional human translations were included as a reference, and they reached the winning cluster in only six of the 15 pairs where they were available. Gemini matched or surpassed them in 10 of its 16. The paper is careful about this, noting it may partly reflect the stylistic preferences of the annotators, but the direction is clear enough.

Two things the paper found that contradict the popular advice

First, speech content was the hardest domain to translate and literary text was the easiest. That is the reverse of the common claim that literary translation is where machines fall down and where one particular model pulls ahead.

Second, running the same evaluation with automatic metrics instead of human graders produced a different winner entirely. A specialised system topped the automatic ranking on all but one pair, then dropped when people read the output. Any translation benchmark scored by software rather than by linguists should be read with that in mind.

Why the Rankings You Find All Disagree

Search this topic and you will get a dozen confident, incompatible answers. One page crowns a tool you have never heard of. Another puts a model at the bottom that a third page puts at the top. There are three reasons for that, and none of them is that translation quality is unknowable.

The first is that the pages are written by sellers. The lists currently ranking for this query are published by localisation platforms and translation companies, and they place their own product in first position. That is not a hidden conflict, it is the business model of the content.

The second is that the measurements are not comparable. Some use BLEU, a decades-old string-overlap score that penalises a good translation for choosing a different word than the reference. Some use an in-house composite with no published method. At least one popular "benchmark" scores translations by checking whether the output contains particular strings, which measures keyword presence rather than whether the sentence is any good. Numbers produced that way should not be compared with numbers from a hand-graded evaluation, and they routinely are.

The third is timing. The rigorous evaluation is annual. Frontier models now ship every few weeks. WMT25 tested Gemini 2.5 Pro, GPT-4.1 and Claude 4, and every one of those has been superseded since. This is the honest limit on what anyone can tell you: the best available evidence describes model families and general patterns, not the exact build you will open this afternoon.

Best AI for Translation by Language Pair

Language pair changes the answer more than any other factor, and it is the variable most roundups skip.

For major European pairs, the frontier models and the dedicated engines are close enough that workflow matters more than raw quality. DeepL built its reputation here and many translators still prefer its phrasing in German, French and Spanish. One thing to know before you trust a number on this: DeepL does not appear anywhere in the 59 pages of the WMT25 findings. The three commercial services the organisers tested are anonymised as ONLINE-B, ONLINE-G and ONLINE-W, and the paper never says which company is which. Anyone quoting DeepL a WMT ranking has invented it.

For Chinese, Japanese and Korean, the large general models pull ahead of the older dedicated engines, mostly because these pairs need contextual judgement rather than word substitution. Give the model the surrounding context and tell it who the reader is.

For low-resource languages, expect much weaker results from everything. The WMT25 authors state directly that translation into low-resource languages remains a challenge, and the scores bear it out. If you need Quechua, Bhojpuri or a minority European language for anything that matters, budget for a human check.

Translating Documents and PDFs

Accuracy is not the constraint here. Formatting is. A model that translates a PDF beautifully and returns it as an unstructured wall of text has not saved you any work.

One number from WMT25 makes the point. The organisers measured how often each model returned a document-level translation with the paragraph structure intact. GPT-4.1 managed it 97.1% of the time and Gemini 2.5 Pro 95.1%, but Claude 4 managed only 67.5% and DeepSeek V3 just 57.2%. Where a model failed, the organisers had to break the document into paragraphs and translate them one at a time, which is exactly the tedious workaround you are trying to avoid.

In practice, paste the whole document rather than sections, tell the model to preserve headings and lists, and check the structure before you check the wording. Our guide to working with PDFs in AI tools covers the extraction side in more depth.

The Best Free AI for Translation

Free tiers move constantly, so treat any specific cap you read anywhere, including here, as something to confirm on the day. The shape of the market is stable enough to describe.

Google Translate covers far more languages than anything else, having added 110 in a single 2024 expansion, and that breadth is why it is still the right tool for skimming a foreign web page. It is not unlimited, though, whatever you may have read. Google caps pasted text at 5,000 characters at a time, and documents at 10 MB or 300 pages for PDFs. DeepL's free tier is more restrictive but well regarded for European pairs. The free tiers of the big chat assistants translate very well and are limited by daily message allowances rather than by word count, which suits occasional use and not bulk work. On the paid side, costs vary by more than an order of magnitude between models, and we track them on our AI pricing comparison.

One cost detail from WMT25 is worth knowing before you assume the best model is the sensible default. Collecting the evaluation translations cost $250.80 in Gemini 2.5 Pro API charges against $31.70 for GPT-4.1, roughly eight times as much. The organisers note the reason: they left Gemini's reasoning mode switched on, which produced eight times the output tokens. Quality at that setting is not free, and for a paragraph the difference is irrelevant while for a book it is not.

Translation on a Mac and iPhone

Apple's approach is the one structurally different option on this list. No other comparison of translation tools seems to mention it at all.

Live Translation on AirPods lets you hear someone speaking another language rendered into yours, hands free. It works with AirPods Pro 3, AirPods Pro 2, AirPods 4 with Active Noise Cancellation and AirPods Max 2, and it is also built into Messages, FaceTime and the Phone app. According to Apple's support documentation, once you have downloaded the languages, all processing happens on your iPhone and the conversation data never leaves it.

The limits are real and worth stating clearly. It covers nine languages, and the regional variants matter: English (US and UK), French (France), German, Italian, Japanese, Korean, Portuguese (Brazil), Spanish (Spain) and Mandarin Chinese in both scripts. Canadian French, European Portuguese and Latin American Spanish are not on Apple's list. It also requires an iPhone 15 Pro or later running iOS 26 with Apple Intelligence switched on. And despite how it is often described, the AirPods version is not a Mac feature. It routes through the iPhone, so a Mac alone will not do it, although Live Translation in Messages does work on a Mac.

On the Mac itself, Apple Intelligence needs Apple silicon and macOS Sequoia 15.1 or later, and supports 16 languages. It is fine for short passages and system-level rewriting. For anything longer or more demanding you will want a frontier model, which is the argument for keeping several available rather than committing to one. We compare the options in Apple Intelligence alternatives for Mac.

What Not to Paste Into a Translator

Translation is the task where people are most likely to hand over something sensitive without thinking: a contract, a medical letter, an internal memo, a document belonging to a client who never agreed to it.

The rule is simple enough. Consumer free tiers are the least suitable place for confidential text, because retention and training policies vary and change. Paid business tiers and API access generally come with commitments that inputs are not used for training, but that is a promise you should read rather than assume. On-device translation, which is what Apple's version does, is the only option where the question does not arise at all. For anything covered by legal privilege or medical confidentiality, check what your obligations actually permit before the text leaves your machine. Our guide on using AI without giving up your privacy goes through the settings worth changing.

How to Get a Better Translation From Any Model

The difference between a mediocre AI translation and a good one is usually the instructions, not the model. WMT25 also found that state-of-the-art systems still struggle with non-standard input and linguistic complexity, which is another way of saying that context helps them more than it should have to.

1. Name both languages explicitly, including the variant. Portuguese for Brazil is not Portuguese for Portugal.
2. Say who will read it and where it will appear. A support email and a legal notice need different registers.
3. Supply a short glossary for names, product terms and anything with a house translation.
4. Paste the whole text rather than fragments, so the model can use the surrounding context.
5. Ask it to flag ambiguities instead of silently choosing. Good models will tell you where the source was unclear.
6. Have a native speaker read anything that will be published, signed or relied on.

Video and Live Translation Are a Different Category

Searches for video dubbing, subtitling and live meeting translation land on this topic constantly, so it is worth drawing the boundary. Those jobs involve speech recognition, timing and often voice synthesis, and the tools that do them well are built specifically for it rather than being general models with a translation prompt. Judging them by text translation quality will point you at the wrong product. WMT25's finding that speech was its hardest domain, largely because of speech recognition errors upstream, is a useful warning about how much can go wrong before translation even begins.

The Verdict

If you want one name: Gemini, on the strength of the only evaluation in this field that pays linguists to read the output. If you want the useful answer, it is four names, not one. Gemini for long documents. Claude or GPT when tone carries the meaning. Apple's on-device translation when the text is sensitive or you have no signal. Google Translate when you only need the gist.

What you should not do is trust a ranking published by a company selling the tool it ranked first, or a benchmark that will not tell you how it scored anything. The evidence in this field is thinner and older than the confident lists suggest, and knowing that is worth more than another table of invented percentages. Keep more than one model within reach, because the right answer changes with the language pair and the job.

Frequently Asked Questions