Microsoft Translator

The apps exist, work, and cost nothing. There are iOS, Android, Amazon and Windows builds, offline language packs, camera translation of signs and menus, an Outlook add-in and a browser-based conversation mode. Microsoft also maintains an education pitch around live captioning and cross-language communication with parents, which schools have used for years.

What there is not is momentum. The consumer marketing site carries a last-modified date of April 2024, and the feature set it describes is roughly the one it described then. Compare that with Google Translate, which in the same window moved to Gemini models, launched live headphone translation across 70-plus languages and added a language-practice mode. Microsoft's consumer translator is being maintained, not developed.

The engine is now sold as Azure Translator inside Microsoft's Foundry Tools reorganisation, and that is where the interesting work landed. The 2026 text-translation API versions let a caller choose, per request, between classic neural machine translation and a generative large language model — a genuinely useful control, because the two fail differently. Neural MT is fast, cheap and predictable; a generative model handles context, tone and ambiguity better and costs more. Being able to route a product-catalogue string one way and a customer email the other, in the same integration, is the kind of feature that only matters at volume and matters a lot there.

Around it sits document translation that preserves formatting, and Custom Translator, which lets an organisation tune the engine on its own bilingual material so that internal terminology survives.

More than 100 languages, which is enough for almost any commercial deployment and short of Google's 249. Microsoft does not market quality claims as loudly as DeepL does, and in practice nobody chooses this service by running a bake-off. They choose it because the invoice already says Microsoft.

It is fair to raise, because Microsoft has form. The Translator Hub was retired in 2019 and the V2 API in 2021, both with migration paths and both painful for the teams affected. The current v3 and 2026-dated API versions are the supported surface; anyone building on it should read the versioning notes rather than assume permanence, and should keep an eye on the consumer apps in particular, which are exactly the sort of unmonetised satellite Microsoft has retired before.

The most-used translation feature Microsoft ships is not the app at all. It is the translate button in Outlook, the translated live captions in Teams meetings, the automatic page translation in Edge and the language tools in Word. None of that is a separate purchase, and for a large share of office workers it is the only machine translation they will ever knowingly use.

That is also the strongest argument for the service. Translation quality differences between the major engines are real but modest for comprehension tasks; the difference between a tool that is already deployed, compliant and licensed and one that requires a new vendor is not modest at all.

The feature that closes deals is not quality, it is terminology. Custom Translator lets an organisation train the engine on its own bilingual material — previously translated manuals, product catalogues, support articles — so that internal names, part numbers and house phrasing survive the round trip. A manufacturer with forty years of translated documentation can make the machine sound like its own technical writers.

That is a different value proposition from anything a consumer translator offers, and it is why the Azure side of this product has customers who would never evaluate DeepL. It is also work: you need clean parallel data, someone to manage the training, and a reason for the investment to pay back.

Three, based on the documented history. Check which API version you are targeting and read the deprecation policy — Microsoft retired the Translator Hub in 2019 and the V2 API in 2021, both with notice and both disruptive. Check whether your workload should route to neural or generative models per request now that the choice exists, because the cost difference at volume is substantial. And check where the free tier ends: Azure's allowances change, and a proof of concept that fits inside the free allocation can become a real invoice the month it ships.

The one consumer-adjacent area where Microsoft has kept investing is education, and it is worth knowing about because nothing else covers it as well. Live captioning of a lesson, translated in real time for students who do not share the teacher's language, plus multi-device conversation for parent evenings, is a genuine accessibility tool rather than a demo. Schools have been running it for years, it costs nothing, and it works on the hardware they already have.

That is also a reminder of why the consumer apps still exist despite the neglect. They are not a business line; they are a demonstration surface and a public good attached to a service that makes its money elsewhere.

If you run on Microsoft, use Microsoft Translator and spend the saved procurement effort somewhere it matters. If you are a developer who needs to mix cheap bulk translation with higher-quality generative passes in one API, the Azure version is currently the most flexible of the big three. If you are an individual choosing a translation app on the merits, download Google Translate first and DeepL second — this one earns its place through integration, not through the sentence it produces.