How Does Proxylang's Translation Score Against DeepL and Google Translate?

People ask us one question more than any other: is the translation good? So we tested it against the two engines most website translation tools run on, and we’re publishing all of it, including the pairs where we lose.

The scores

	Proxylang	DeepL	Google Translate
Average, all 280 sentences	4.72	4.61	4.20
Scored a perfect 5 (Claude Opus 5.5)	74.3%	63.6%	42.1%
No errors found (GEMBA-MQM)	46.1%	40.4%	30.7%

If you’re comparing website tools: Weglot’s help center says its automatic translations come from DeepL whenever DeepL supports the pair. So for Korean, Japanese and English, the DeepL column is Weglot’s starting point.

By language pair

Pair	Proxylang	DeepL	Google
Korean → Japanese	4.69	4.46	3.10
Japanese → Korean	4.84	4.71	3.79
Korean → English	4.76	4.38	4.21
Japanese → English	4.78	4.55	4.34
English → Japanese	4.66	4.65	4.61
English → Korean	4.72	4.85	4.54
English → German	4.61	4.64	4.79

The biggest gaps are between Korean and Japanese. Google Translate often reads a Korean word as the wrong Chinese characters. 초인종 (doorbell) came out as 超人種, “superhuman race”. The Japanese looks real, so nobody who doesn’t read Japanese would catch it. We collected more real examples here.

Where we lose

AI judges don’t give the exact same score twice. We scored the same, unchanged DeepL and Google translations three times, and their averages moved by up to 0.09 between runs. So read small gaps as ties:

• English → Japanese: 4.66 vs DeepL’s 4.65. A tie.
• English → German: 4.61 vs DeepL’s 4.64, a tie. Google is best here at 4.79.
• English → Korean: DeepL 4.85, us 4.72. The gap is 0.13, a bit more than the scores move on their own, so DeepL is slightly ahead. The points we lost were small rewordings, not wrong meanings.

How we tested

The sentences. 40 per language pair from FLORES-200, a public test set of news and science sentences. Each one comes with a translation by a professional translator.

The systems. Proxylang with its normal translation setup, the DeepL API on its quality-optimized model, and the Google Cloud Translation API. We turned off everything a real website gives Proxylang: no glossary, no brand names, no page context. That makes this our plain first pass.

The judges. Claude Opus 5.5 (Anthropic) and GPT-5.6 Terra Pro (OpenAI) each scored every translation on their own. Neither company makes Proxylang, DeepL or Google Translate. Proxylang runs on a Google model, so we left Google’s models out of the judging.

Two scoring methods.

1. A 1 to 5 scale. The judge sees the source, the professional translation and all three candidates, with no names and in a new random order for every sentence. Both judges ranked the three systems in the same order: Claude gave 4.71, 4.58 and 4.13, and GPT-5.6 gave 4.74, 4.64 and 4.27.
2. GEMBA-MQM, a published method from WMT 2023. The judge sees one translation at a time, with no reference and nothing to compare against, and lists every error as critical, major or minor. Claude Opus 5.5 checked all 840 translations this way, before the Japanese update below. Proxylang had the fewest errors.

What changed since our first run. We first scored these translations on October 14 with a smaller open model, Qwen 3.6 35B. It put us level with DeepL: 4.53 against 4.54. People asked us which model did the judging, and a small model is a weak answer, so we re-scored the same translations with two stronger judges. All three systems scored higher, and the order changed in our favor. The same day, we tightened our English to Japanese instructions and re-tested that pair. It went from 4.60 to 4.66.

Limits. 40 sentences per pair is small, and scores move by up to 0.09 between runs on their own. AI judges are not human translators. And news sentences are not website copy.

On your website, it keeps getting better

This test scores our first attempt at a bare sentence. On your site, Proxylang has more to work with:

• Your glossary and never-translate list. Product names come out the same in every language.
• The whole page. Each sentence is translated knowing what the page is about.
• Re-checks after launch. Proxylang keeps re-reading the translations live on your site, against the rest of your site and your glossary. When a line reads stiff, a name got translated, or the meaning slipped, it puts in a better version. Re-checks are free, don’t use your word allowance, and never overwrite your own edits. How it works.

Links and bold text inside a sentence also move with the words in Japanese and Korean, so a link never ends up half on a word. See examples.

Coba di situs Anda sendiri

The free trial needs no signup and covers 2,000 words in one language. Paste this into Claude Code, Cursor or any AI coding tool:

Your agent sets up the trial for you. Or open proxylang.dev/start and type your site’s address.

Want the full write-up, with every example sentence? It’s on Proxylang Answers.