For most apps the best translation API is a hosted cloud engine billed per character, chosen by running a blind test of two or three on your own text. If your text can't leave your own servers, pick an open-source engine you host yourself or an on-premise licence instead.
Start here: which kind of engine fits you
Before you compare a single vendor, decide which shape of product you need. Find your row, then read only the entries that match it.
| If this sounds like you | Look at this type | How the bill grows |
|---|---|---|
| You want translation live in your app this week | Hosted cloud API | With every character you send |
| Tone and wording matter more than raw volume | LLM-based endpoint from a cloud provider | With characters, often counted in and out |
| Nothing can leave your servers | Open-source engine you host yourself | With your hardware and staff time, no meter |
| Huge, steady volume or offline devices | On-premise server or offline SDK | A flat licence rather than per character |
That third column is the one that bites later. Per-character billing is gentle at prototype stage and grows in step with your users. A flat licence costs more on day one and less with every word after. Self-hosting has no meter, but someone on your team now owns a server, model updates and whatever breaks at 3am.
Translation engines to put through the test
DeepL API
Translation API for text and whole documents, with separate free and paid endpoints so you can prototype without a card.
Google Cloud Translation
Google's cloud service covering 189 languages, with pre-trained, custom and LLM-based models under one API.
Amazon Translate
If your stack already lives on AWS, this is the machine translation service that slots in beside it for text and documents.
Lara Translate
An AI translator for text, documents, images and audio across 200+ languages, sold with a CAT tool as well as API access.
SYSTRAN Translate API
A translation API pitched at developers who want machine translation built straight into their own applications, with a live demo to try first.
Lingvanex
Teams that need translation offline or behind their own firewall will find cloud API, on-premise server and mobile SDK options here.
Tilde Machine Translation
Tilde's machine translation product handles texts and documents, and there's an enterprise tier for organisations translating at scale.
LibreTranslate
Open-source translation API you can download and run yourself, or use hosted if you'd rather not manage a server.
Argos Translate
For Python developers who want offline neural translation, it ships as a library, web app, API and desktop app.
Run a blind test on your own strings
Don't pick on reputation. Language pairs and subject matter swing results too much for anyone's ranking to hold for your product, and the free plans and trials on this list make a fair test cheap.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
- Pull 30 real strings from your product: ten short button labels, ten sentences from your help docs and ten lines of marketing copy with some personality in them.
- Choose the two language pairs that bring in the most revenue, not the easiest ones.
- Send every string through three engines and paste the output into a spreadsheet, one column per engine.
- Shuffle the columns so nobody knows which engine wrote what, and ask a native speaker to mark each line fine, fixable or wrong.
- Count the wrong lines per engine. Then check the mechanics: did placeholders like {name} come back untouched, did HTML tags survive, and do the translated labels still fit your buttons?
The engine with the fewest wrong lines on your hardest pair usually wins. If two are level, let billing and data handling settle it. If one of them offers an LLM model alongside its standard one, run both through the same sheet; idiom and brand voice are the things worth testing that way.
What's moved in the category this year
LLM translation now sits inside the classic cloud APIs. You don't have to choose between a neural engine and a large language model; at least one big cloud sells both under one product. On Google's own pricing page, its Translation LLM is listed as cost equivalent to its standard neural model, so don't assume the LLM route costs more. Check whether a rate counts input and output characters separately.
Translation is also turning up inside AI assistants. One big engine now says it's reachable from Copilot, ChatGPT and Claude through MCP. That's handy for staff, but it doesn't replace an API call inside your checkout, so don't let it decide an engineering choice.
Then there's the warning sign. ModernMT says on its own homepage that it's becoming Lara and will sunset by the end of 2026. Teams built on it are migrating whether they planned to or not.
Do you make translation APIs or machine translation engines? Tool makers can ask to be considered for a featured place on this list.
How the bill works, and how to keep it small
Cloud translation is mostly billed per million characters, so your bill is a straight line from how much text you send. On the two biggest clouds' own pricing pages this month, standard text translation lists at $15 to $20 per million characters once any free allowance is used up. Some vendors sell on-premise, unlimited-volume licences instead, priced by the month rather than the character. Open-source engines cost nothing to licence; at least one also sells a hosted key on a monthly plan if you don't want the server. Prices change, so take today's rate from the vendor's own pricing page before you budget.
To size it, count characters rather than words. A help centre of 2,000 articles averaging 6,000 characters is 12 million characters per language. At $15 to $20 per million, that's roughly $180 to $240 per language for the first pass. After that you only pay for what changed, if you build it that way:
- Cache every result against the source text and target language, so a string is never translated twice.
- Send only new or edited strings on each deploy. Re-translating everything every release is how a small bill becomes a large one.
- Use a glossary or do-not-translate list for product names, feature names and legal terms from day one.
- Keep the vendor behind one function in your code. A price rise or a sunset notice then means changing one file, not a rewrite.
Privacy and the question legal will ask
Someone will ask whether the vendor stores your text or trains on it. Read the data terms for the exact plan you're buying, not the vendor's general marketing, and keep a dated copy. If the answer has to be that nothing leaves your network, the table at the top has already narrowed it: host an open-source engine yourself, or buy an on-premise licence so there's a vendor to call when something breaks. Both move cost from the invoice to your own team; that's the trade, and it's often worth it for health, legal or government text.
If your output is headed for video rather than an app, timing and line length change the job; our guide to subtitle translators covers that. And if you've already used one of the big engines and want like-for-like options, see these DeepL alternatives.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

