Which translator is the most accurate?

What the independent studies measured

The studies here tested particular languages and kinds of text, not every task you might need a translator for. Several used hospital discharge instructions. That's worth keeping in mind before you apply their results to an essay or a store page.

Kong et al., in BMJ Quality & Safety in 2025, compared ChatGPT-4 with Google Translate. The share of sentences translated accurately was:

• Spanish: 97% for ChatGPT-4, 96% for Google.
• Chinese: 95% versus 90%.
• Russian: 89% versus 80%.

For Chinese, 24% of GPT-4 instruction sets contained at least one error, compared with 56% for Google. The authors still advised against machine translation for high-stakes communication.

No blanket winner here. Those results tell you about those languages, those instructions and the models tested.

Where Google Translate is strong and where it slips

Google Translate covers almost 250 languages, according to Google in April 2026. If you're choosing by language coverage, that's a concrete reason to consider it.

Accuracy needs a closer look. Khoong et al., in JAMA Internal Medicine in 2019, found that Google got 92% of Spanish and 81% of Chinese discharge-instruction sentences right. But 2% of the Spanish sentences and 8% of the Chinese sentences could have caused harm.

The gap between languages was wider in Taira et al., 2021: Google kept the meaning in 94% of Spanish statements, but only 55% of Armenian ones.

So, is Google accurate? For your decision, "in which language?" matters more than a single overall score.

ChatGPT and other AI models

ChatGPT-4 beat Google in the Kong study, but that doesn't make ChatGPT the best choice for every language.

Brewster et al., in Pediatrics in 2024, found clinically significant errors in Haitian Creole in 33.3% of ChatGPT translations, 23.3% of Google Translate translations and 8.3% of professional translations.

And the AI comparison goes beyond ChatGPT. In WMT24, the yearly machine translation benchmark run by researchers, Claude 3.5 Sonnet ranked best in 9 of 11 language pairs in human evaluation. At WMT25, Gemini 2.5 Pro was in the top group for 14 language pairs.

Those results belong to the named models. Don't read an older ChatGPT result as a test of every version you can use in 2026.

What DeepL's own tests claim

DeepL says it won 89% of head-to-head groups, or 66 of 74, in blind tests it ran in June 2026 across 18 language pairs, including 18 of 18 against Google Translate.

These were DeepL's own tests. They weren't independent.

You can consider that claim without treating it as the same kind of evidence as an independent study. It also doesn't settle the choice for a language pair or type of text outside what was tested.

How to pick for your language

Start with the language you're translating from and into, then look for evidence that matches your text. A result for Spanish discharge instructions doesn't answer how a tool will handle your Armenian document or Dutch store copy.

For important text, compare translations of your actual material and have someone qualified in the language check the meaning. For high-stakes communication, follow the Kong authors' advice against machine translation.

Want to try tools before paying? As of November 2026, DeepL's free web translator allows 50,000 characters a month, and ChatGPT's free plan allows unlimited everyday text chats. Those are access limits, not accuracy scores.

Text with code in it

With website text, check the code as well as the words.

Translinga tested the DeepL API from English to Dutch on October 27, 2026. DeepL changed the Shopify filter | money to | geld, after which Shopify prints the raw amount in cents. It also translated "Save" in a combat menu as "Opslaan," meaning saving a file, even with context supplied. On a billing page, "Charge" became "Opladen," meaning charging a battery.

That's a different check from asking whether a sentence sounds right.

Translinga, made by Bitgate for Shopify stores, WordPress sites and apps, removes Liquid tags, HTML and placeholders before translation, checks replies tag by tag, and retries strings with missing tags. You can add a $context note to a JSON file, and key names count as context too.

It's in early access through the Get early access button on translinga.com. Fast costs $2.50 per million characters; Quality costs $5.00 per million. Spaces count, once per target language. Prepaid credit starts at $5 and never expires; the first $1 is free (about 400,000 characters on Fast), with no card needed for a key.