The best AI translators in 2026

What the benchmarks say

The benchmark results give you a shortlist, but they don't name the best tool for your text. Pay attention to the model version and the language pairs being tested, not just the company name.

In WMT24, the yearly research benchmark for machine translation, Claude 3.5 Sonnet ranked best in 9 of 11 language pairs in human evaluation. Human reference translations were in the winning group for 7 of 11.

WMT25 put Gemini 2.5 Pro in the top group for 14 language pairs, while human references made the winning group for 6 of 15. Those are results for those versions. They aren't a score for every model you can open today.

There's another approach worth noting: Intento's State of Translation Automation 2025 compared 46 engines and LLMs, and a multi-agent setup came out best in 9 of 11 language pairs.

No single winner, then. Use these results to pick what to test, not to skip the test.

General chatbots: ChatGPT, Claude and Gemini

If you're choosing among ChatGPT, Claude and Gemini, Claude and Gemini have the benchmark results above. ChatGPT also has a narrower result worth looking at if your material includes medical instructions.

Kong et al. 2025, published in BMJ Quality & Safety, found ChatGPT-4 ahead of Google Translate on Chinese and Russian medical instructions, and about even on Spanish. The share of accurate sentences was 97%, 95% and 89% for ChatGPT-4, versus 96%, 90% and 80% for Google Translate.

That's evidence about those medical instructions and those models, not a blanket recommendation for medical translation.

What about cost? As of November 2026, ChatGPT Plus costs $20 a month. The OpenAI API lists gpt-6.1-sol at $2 per million input tokens and $10 per million output tokens.

Keep the subscription and API prices separate when you compare your options. And don't read the ChatGPT-4 study as an evaluation of gpt-6.1-sol.

Dedicated translators: DeepL and Google Translate

DeepL has a strong result from its own testing: in June 2026 blind tests, it says it won 89% of head-to-head matchups against 7 competitors, including Google Translate and GPT-5.5.

Keep the source attached. These were vendor-run tests, not WMT results.

For a free starting point, DeepL's web translator includes 50,000 characters a month. Google Translate is free, and its Cloud API costs $20 per million characters after 500,000 free a month.

If you're choosing between them, try the same source text in both (not a different sample for each). For a document, include the passages you're least sure how to translate, rather than testing only the opening paragraph.

AI translation for websites and apps

For a website or app, test the strings with their markup and placeholders included, not just the words you'd paste into a chat window.

Translinga, made by Bitgate, is a translation API for Shopify stores, WordPress sites and apps. It takes Liquid tags, HTML and placeholders out before translation, checks every reply tag by tag, and retries any string that comes back with a tag missing.

You can also put a $context note in a JSON file to explain what the strings are for. Key names count as context too.

Translinga's Fast tier costs $2.50 per million characters; Quality costs $5.00 per million. Spaces count, and the character count applies once per target language. Credit is prepaid from $5 and never expires.

It's in early access. You request a key through the Get early access button on translinga.com, with no card needed. The first $1 is free, about 400,000 characters on Fast.

How to test before you commit

Start with your own material. Pick text you intend to publish, a passage from your document or strings from your product, and give each tool the same source and the same context.

What would make you reject a translation? Write that down before comparing the replies. Check the meaning, the terms you want kept consistent and any wording you don't want changed.

For product strings, check the tags and placeholders too. If you're testing the JSON context option, include your actual key names and a note about what the strings are for, then review the output in that context.

Keep cost in the test. Record the characters or tokens you're billed for, using the unit each service charges by.