AI Translators Compared: DeepL to Local

Which AI translator is right depends less on quality than on what your text contains. For harmless texts, the free DeepL Translator and Google Translate both work. In a test with the same client email, DeepL, Google and the local language model Qwen agreed on Inbetriebnahme, Schaltschrank and Abnahme but not on Abschlagsrechnung; given a glossary, Qwen used the required term in both runs. Once client data or anything confidential is involved, the free versions drop out: DeepL's free version forbids personal data, and Google analyzes Translate usage to improve it. Then you need a paid plan with a data processing agreement (DeepL Pro, Google Cloud Translation, Claude or ChatGPT on a business plan) or a local model where the text never leaves your network. All prices and privacy details as of October 30, 2026.

So you can weigh my view: I pay for Claude and ChatGPT but work almost exclusively with Claude, because my workforce runs in Claude Code. DeepL and Google were tested on October 30, 2026, the local model on November 1, 2026. Claude and ChatGPT were not part of this test; what follows about them comes from the providers' documentation. This isn't legal advice, just the operator's perspective I use for my own business in Germany, so it's about the GDPR and the EU. More basics are on my AI Basics page.

AI translators at a glance: four routes

Claude and ChatGPT list prices in US dollars on their US pages, Anthropic excluding tax; from Germany, claude.com shows euro prices, for example Pro at €18 a month or €15 billed annually. All Claude plans are listed in Claude Pricing and Plans Explained.

The test: the same client email in three tools

The test text is a made-up German business email with typical stumbling blocks: technical terms, an idiom and the formal “Sie” form of address.

Sehr geehrter Herr Hughes,

vielen Dank für Ihre Rückmeldung zu unserem Angebot vom 14. Oktober. Die Inbetriebnahme der Verpackungsanlage verschiebt sich leider um eine Woche, weil das Ersatzteil für den Schaltschrank erst am 6. November eintrifft. Das ist ärgerlich, aber kein Beinbruch: Ihr Liefertermin an den Endkunden bleibt davon unberührt. Die Abschlagsrechnung über 30 Prozent stellen wir deshalb erst nach der Abnahme. Ich halte Sie auf dem Laufenden, und bei Fragen rufen Sie mich gern direkt an.

Mit freundlichen Grüßen

Setup for DeepL and Google: one automated browser run per tool via Playwright, not signed in, no glossary, default settings (cookie prompt declined on Google), target English (DeepL: “English (American)”). The setup for the local model is in its test block below. Outputs verbatim.

DeepL, free web translator (October 30, 2026):

Dear Mr. Hughes,

Thank you very much for your response to our quote dated October 14. Unfortunately, the commissioning of the packaging system will be delayed by one week because the replacement part for the control cabinet will not arrive until November 6. This is frustrating, but it’s not the end of the world: Your delivery date to the end customer will not be affected. We will therefore issue the 30 percent advance invoice only after acceptance. I will keep you updated, and please feel free to call me directly if you have any questions.

Best regards,

Google Translate, web (October 30, 2026, “Erweitert” (Advanced) preselected):

Dear Mr Hughes,

Thank you for your feedback on our offer from October 14th. Unfortunately, the commissioning of the packaging system has been postponed by a week because the spare part for the control cabinet will not arrive until November 6th. This is annoying, but it's not a problem: your delivery date to the end customer remains unaffected. We therefore only issue the advance invoice for 30 percent after acceptance. I will keep you updated and if you have any questions, please call me directly.

Best regards

Local: Qwen3.8-27B on an Nvidia DGX Spark (November 1, 2026)

Setup: Qwen3.8-27B as an NVFP4 checkpoint under the SGLang serving framework, called through the OpenAI-compatible API directly on the device, temperature and other values at the server defaults. Prompt (in German): “Übersetze diese E-Mail ins amerikanische Englisch. Behalte den förmlichen Ton bei.” (“Translate this email into American English. Keep the formal tone.”), followed by the email. Second run in the same chat with the glossary: “Glossar: Abschlagsrechnung = progress invoice. Übersetze die E-Mail noch einmal und halte dich an das Glossar.” (“Glossary: … Translate the email again and stick to the glossary.”) Both runs once with thinking mode off and once with thinking mode at the server default, one pass each. Time is from sending to the finished answer.

Thinking off, no glossary (3.7 seconds):

Dear Mr. Hughes,

Thank you for your response to our quotation dated October 14.

Unfortunately, the commissioning of the packaging line will be delayed by one week because the replacement part for the control cabinet will only arrive on November 6. While this is regrettable, it does not affect your delivery deadline to the end customer. Consequently, we will issue the 30% down payment invoice only after final acceptance.

I will keep you informed of any updates, and please feel free to contact me directly with any questions.

Sincerely,

Thinking off, with glossary (2.4 seconds):

Dear Mr. Hughes,

Thank you for your feedback regarding our offer of October 14. Unfortunately, the commissioning of the packaging line will be delayed by one week, as the replacement part for the control cabinet will only arrive on November 6. While this is regrettable, it is not a major issue: your delivery deadline to the end customer remains unaffected. Therefore, we will issue the progress invoice for 30% only after the final acceptance.

I will keep you informed of any updates, and please feel free to contact me directly with any questions.

Sincerely,

Thinking on, no glossary (38.5 seconds):

Dear Mr. Hughes,

Thank you for your response regarding our proposal dated October 14. Unfortunately, the commissioning of the packaging line will be delayed by one week, as the replacement part for the control cabinet will not arrive until November 6. Although this is regrettable, your delivery date to the end customer will not be affected. The invoice for the 30% progress payment will therefore be issued only after acceptance. I will keep you informed, and if you have any questions, please feel free to call me directly.

Sincerely,

Thinking on, with glossary (21.2 seconds):

Dear Mr. Hughes,

Thank you for your reply regarding our proposal dated October 14. Unfortunately, the commissioning of the packaging line will be delayed by one week because the replacement part for the control cabinet will not arrive until November 6. While this is regrettable, your delivery date to the end customer will remain unaffected. The progress invoice for 30 percent will therefore be issued only after acceptance. I will keep you informed, and if you have any questions, please feel free to call me directly.

Sincerely,

What came out at the check points

All of them agree on the technical terms Inbetriebnahme, Schaltschrank and Abnahme. They differ on everyday words: DeepL turns Angebot into “quote”, Google into “offer”, Qwen into “quotation”, “offer” or “proposal” depending on the run. Only DeepL keeps the idiom figurative. Google turns “kein Beinbruch” into a plain “not a problem”; Qwen drops it entirely in three of four runs and turns “ärgerlich” (annoying) into a formal “regrettable”. That matches the “formal tone” instruction, which only Qwen received.

The Abschlagsrechnung shows what a glossary is for. DeepL and Google make it an “advance invoice”, Qwen without thinking a “down payment invoice”; both sound like prepayment, yet in the same sentence the invoice comes only after acceptance. Qwen with thinking gets closer with “progress payment”. With the glossary, both runs said exactly “progress invoice”. Without thinking, though, “call me” became a general “contact me”; with thinking the call stayed, but the answer took nine to ten times as long.

Rated for glossary adherence, errors and tone:

Each setting ran once. A second pass can produce different words, especially for everyday terms like Angebot. Whatever you want fixed belongs in the glossary, not in the hope that the tool picks the same word next time.

DeepL Free or Pro: the privacy difference

DeepL's privacy policy (October 2026) separates the two versions clearly. For the free translator, DeepL processes your content and translations “for a limited period of time to train and improve our neural networks and algorithms”, and its terms say you may not use it for “texts containing personal data of any kind.” A client email with a name in it is already out.

With DeepL Pro, texts “will not be permanently stored” according to DeepL, are deleted after delivery, and “will not be used to improve the quality of our services”; for personal data, the Pro terms provide for a data processing agreement. If you want a DeepL alternative because the free version feels too risky, the first alternative is often DeepL Pro itself.

Google Translate: web interface versus Cloud API

The free Google Translate falls under Google's general privacy policy (effective October 16, 2026), which gives this example: “We analyze usage of Google Translate to improve translation quality and increase the availability of Translate in more languages.” That is not the route for client data.

The paid Cloud Translation API is different: according to Google, text is held only briefly in memory, not used for training, and covered by the Cloud Data Processing Addendum. It's a developer building block, not a box to paste into, but the more privacy-friendly Google route for translation inside your own software.

Claude and ChatGPT as translators

The difference from DeepL and Google is the instruction: you can tell a language model who the text is for, which terms are fixed and which tone you want. The glossary run with the local Qwen above shows the effect; Claude and ChatGPT were not part of this test. The general comparison of both tools is in Claude vs ChatGPT 2026: Compared. Proofreading stays a must, see AI Hallucinations: Why AI Makes Things Up.

On privacy, the plan decides. On Claude Free, Pro, Max and Ultra, a toggle decides whether Anthropic trains on your chats; Anthropic lists this as opt-out, and if allowed, data stays in the training pipeline for up to five years. On ChatGPT's personal plans, sharing is on by default according to OpenAI and can be turned off. Both providers say their business plans and APIs don't train on your inputs by default, and both offer a data processing agreement there, at Anthropic automatically part of the commercial terms. Client emails belong in the business plan. The settings per Claude plan in detail are in Claude Privacy and GDPR: What Applies.

Translating locally with Ollama and Qwen

If a text must not leave your network, the model runs on your hardware. I run Qwen3.8-27B under SGLang on an Nvidia DGX Spark, the model from the test above; setup and measurements are in Qwen 3.8 Locally: DGX Spark Experience, the overview is in Local AI: What Actually Works in 2026. The model is licensed under Apache 2.0; in the Ollama library the 27B version is 18 GB. How to start Qwen 3.8 with Ollama is in Installing Qwen 3.8 with Ollama. According to Ollama, the provider sees neither your prompts nor your data when you run locally.

In the test, the translation without thinking mode arrived after 2.4 to 3.7 seconds and kept to the glossary. With thinking mode the call stayed and the Abschlagsrechnung came out better even without a glossary, but the answer took 21 to 38 seconds. The times apply to this hardware and one pass each.

The price is effort for hardware, setup and upkeep. Getting from a Mac to your own server is covered in Setting Up Local AI: Ollama to SGLang. Local doesn't automatically mean compliant either; the obligations that remain are in Local AI and Privacy: Cloud or Server?.

Which AI translator for which text

• Menu, product text, public web page: free DeepL or Google, proofread anyway.
• Client email, quote, application: a plan without training and with a data processing agreement: DeepL Pro, Claude Team or ChatGPT Business, or a local model.
• Contracts, HR files, health data: a local model or a human professional translator.
• Video subtitles: a separate topic, see Auto-Generate Subtitles: A Guide.

Put the rule in an AI Usage Policy for Small Teams, or the client email lands in the free translator anyway: Shadow AI on a small scale.

Frequently asked questions

Which AI translator is the best?

None across the board. In the tests on October 30 and November 1, 2026, DeepL, Google and the local Qwen agreed on Inbetriebnahme, Schaltschrank and Abnahme. DeepL stayed closest to the idiom; Qwen, as a language model, followed the glossary it was given but wrote more formally. For confidential texts, what counts is whether the provider stores them or trains on them.

What is a good DeepL alternative?

For privacy, often DeepL Pro itself, which according to DeepL neither trains on nor permanently stores your texts. Beyond that: Claude or ChatGPT on a business plan for translations with instructions, or a local model like Qwen if nothing may leave your network.

Can I paste client emails into the free DeepL Translator?

Not under DeepL's terms once personal data is involved, and a name in the greeting is enough. DeepL points to Pro with a data processing agreement. Not legal advice.

Do Claude and ChatGPT translate better than DeepL?

Not measured here; Claude and ChatGPT were not part of the test. What was measured is another language model, the local Qwen3.8: it used the glossary term in both runs, but picked its own words without a glossary and mostly dropped the idiom. Whether Claude and ChatGPT behave similarly is quickest to find out with your own run, your own text and your own glossary.

What to do next

Run a harmless email from your daily work through two tools before subscribing, and set up a glossary for your key terms. For confidential texts, check the plan first, then the quality. How others set up their AI tools with privacy in mind is something we discuss in my English-language community AI Employee Framework.