Machine translation in 2026 is a different field than five years ago. Modern LLMs are often better translators than dedicated translation services — but only if you know how to prompt them. For some workflows, DeepL still wins. For others, Google Translate is fine. The question is which workflow you have.
DeepL: still the leader for European languages
DeepL has been beating Google Translate on European-language pairs for years and the lead persists. German↔English, French↔English, Italian↔English, Spanish↔English — DeepL produces more natural output, handles idioms better, and gets formal/informal register right more often. For business writing in EU languages, it's still the default.
DeepL's secret sauce is training data quality and a focus on a smaller language set. They optimize hard for the languages they support rather than spreading effort across 130+ languages.
Use DeepL when: you're translating between European languages, you need a polished business-grade translation in seconds, you want a UI optimized for translation work (not a generic AI chat). The Mac/Windows desktop apps are still better than the web interface.
Weakness: Asian languages are weaker than the European ones. DeepL's Chinese, Japanese, Korean translations are good, but no longer state-of-the-art. The pricing is also high relative to LLM-based alternatives now.
Google Translate: the pragmatic everywhere choice
Google Translate covers 130+ languages, is free for casual use, has the best mobile app, and integrates everywhere. For low-stakes, daily use — "what does this Vietnamese restaurant menu say," "translate this Japanese tweet" — Google is the right answer.
Google has also been quietly improving its translation engine with Gemini-derived models. The 2024-2026 generation is dramatically better than the old NMT system. For long-form content the old "machine translation feel" is mostly gone.
Use Google when: you need broad language coverage, you want a free option, you need it on mobile, or you need OCR + translate + speak workflow. The Lens camera translate feature is genuinely magical when traveling.
Weakness: registers and tone are unreliable. It will translate a casual tweet into formal business prose. For published content where voice matters, Google's output usually needs human editing.
LLM-direct: the new default for content with context
This is the big shift in 2025-2026: for any translation where you can describe the context, sending the text to Claude, GPT, or Gemini directly with a careful prompt produces better results than DeepL. Not always; but often.
Why: dedicated translation systems treat each sentence in isolation. LLMs can use the surrounding paragraph, your stated audience, your stated register, and your stated brand voice as context. "Translate this tweet for a casual Taiwan audience, keep tech terms in English, don't formalize the tone" is something DeepL can't do.
Use LLM-direct when: you're translating a longer piece (article, doc, book), the audience matters, the brand voice matters, you have terminology you want preserved (like keeping "prompt" in English in Chinese tech writing), or you want to translate to a less-common language pair (English → Vietnamese, Chinese → Indonesian) where dedicated services lag.
A useful prompt structure: "Translate the following [source language] text into [target language]. The audience is [audience description]. The tone should be [tone]. Keep these terms in [original language]: [list]. Do not localize culturally — translate, don't reframe. Output only the translation." Then paste the source.
Specialized tools you might miss
- Lokalise / Crowdin / Phrase — translation management systems for software. Keep a glossary, manage translator workflows, integrate with Git. Use when you're shipping a product in 5+ languages.
- CAT tools (Trados, MemoQ) — for professional translators. Translation memory, terminology databases. Overkill unless translating is your actual job.
- Subtitle Edit / Aegisub — subtitle-specific tools. Different workflow from prose translation; line-length and timing constraints matter.
- Notta / Vatis Tech — translate audio/video by transcribing, translating, and dubbing. Niche but useful.
Chinese-specific considerations
For Chinese, the landscape is more complex:
- DeepL — solid but not best-in-class
- Google Translate — okay, sometimes uneven
- GPT-5 / Claude 4.5 / Gemini 2.5 — best for Chinese↔English with context, especially for long-form
- DeepSeek / Qwen — Chinese-developed models, often better at Chinese-internal nuance (Traditional vs Simplified, Taiwan vs Mainland register, technical terms)
- Baidu Translate / Youdao — old standbys, fine for casual, dated for serious work
For Traditional vs Simplified specifically: don't just rely on character conversion. The vocabulary differs (软件/軟體, 视频/影片, 默认/預設, 信息/訊息). Tell the LLM explicitly which variant you want.
When NOT to use any machine translation
- Legal contracts, medical records, anything where a wrong word has real consequences. Use a certified human translator. Machine output is fine as a draft for the human to edit.
- Brand-critical marketing copy in a new market. Translation isn't the same as transcreation. Pay a local writer.
- Anything where idiomatic correctness matters and you can't verify it. Machine translation hallucinations into unfamiliar languages are very hard to catch.
- Content for elderly readers or cultural contexts you don't know — translation gets the words right but not the register.
Cost reality
For 1M characters of translation:
- DeepL Pro: ~$25/mo subscription tier
- Google Translate API: ~$20
- GPT-5 Mini via API: ~$2-3 (depending on prompt length)
- Claude 4.5 Sonnet via API: ~$5-8
- Free Google Translate UI: $0 (fine for personal use)
LLM-direct translation, in pure API cost terms, has become dramatically cheaper than dedicated services for high-volume use.
Decision tree
- European business writing: DeepL
- Casual everyday, mobile, broad languages: Google Translate
- Long-form content with context, brand voice: Claude 4.5 or GPT-5 with prompt
- Chinese-specific nuance (TW/CN, technical): GPT-5, Claude, or DeepSeek
- Multilingual product UI: Lokalise / Crowdin + LLM-direct for first drafts
- Real-time conversation: Google Translate live, Apple Translate
Next steps
- Learn how to write a translation prompt for LLMs — register, glossary, audience
- For products, look into translation management systems early
- For personal use, install both DeepL and Google Translate apps and use whichever feels right per language
- Test the same paragraph through 3 tools and read the output side-by-side; you'll quickly know which fits your voice