A DeepL alternative for documents

Author: Yuri Vodostoy — Technical translator and founder of Insight Translation · Updated 2026-08-05

HardTrans reads every file in a job before it translates a word, builds a project brief and a glossary out of what it finds, and applies both to every file. DeepL translates what you give it; it does not read the whole job first and derive a project brief and glossary from it. On our German-to-English runs of August 5, 2026, DeepL got two of five engineering passages right, came close on a third, and missed two. Google got none of the five clean, and HardTrans all five. Price is $1.50 per 250-word page, first document up to 10 pages free, no signup.

Is DeepL better than Google Translate?

On these passages, yes, and by a wide margin. DeepL kept the maintenance hierarchy Google collapsed into a tautology, wrote “de-energized” where Google wrote “switched off,” and dodged the role inversion Google walked into on an acceptance clause. Its own failures land on terms whose only correct form is the one printed in the standard.

Nine German passages went through Google Translate first, as a filter. Four came back clean (pressure-vessel spec, manual boilerplate, pump efficiency, fault clearance), so we dropped them; modern MT has industrial boilerplate memorized. Five stayed, the four panels below plus one reserve.

Where DeepL breaks on standards terms

Two passages, and in each one a clause fixes the English, not usage.

Material certificate, EN 10204

Source: «Abnahmeprüfzeugnis 3.1 nach DIN EN 10204. Schmelzen-Nr. 78542. Die chemische Zusammensetzung der Schmelze entspricht den Anforderungen der Norm.»

Tested: August 5, 2026

EN 10204:2004, clause 4.1 (specific inspection documents), names this document “inspection certificate 3.1,” and buyers check certificates against clause numbers. Neither “Acceptance Test Certificate” nor “Inspection test certificate” is in the standard.

The traceability key is the second problem. EN 10168:2004 carries it in field B07 in three languages, Schmelzen-Nr. / No. de coulée / Heat No., and cast no. is equally standard. Melt is not certificate language. Google found “Heat no.,” then wrote “the melt” one sentence later, the drift we see most often across whole documents.

Machine safety, two functions with one name

Source: «Die Maschine ist mit einer Not-Halt-Einrichtung nach EN ISO 13850 ausgerüstet. Zusätzlich trennt die Not-Aus-Einrichtung die Maschine von der Energieversorgung.»

Tested: August 5, 2026

The source names two devices, and zusätzlich shows it. DIN EN 60204-1 separates them by clause, 10.7 for Not-Halt and 10.8 for Not-Aus, and the English edition EN IEC 60204-1:2018 keeps the numbering with “emergency stop devices” and “emergency switching off devices.” Both engines returned one device twice, so a safety section that separates stopping the motion from cutting the power now says the same about both.

Where DeepL pulls ahead of Google

Google breaks in the next two, and it breaks harder.

The maintenance triad

Source: «Die Instandhaltung der Anlage umfasst Wartung, Inspektion und Instandsetzung. Die Instandsetzung defekter Baugruppen erfolgt ausschließlich durch den Hersteller.»

Tested: August 5, 2026

DeepL got the hierarchy right. DIN 31051 sets it out, Instandhaltung on top with Wartung, Inspektion and Instandsetzung under it, and EN 13306 supplies the English, where Wartung is servicing. Google defined maintenance as including maintenance, which no English reader can parse. Where DeepL stayed generic is Baugruppen, and assemblies is the term.

Who runs the test, and where

Source: «Die Dichtheitsprüfung erfolgt vor der Abnahme beim Hersteller. Das Protokoll der Dichtheitsprüfung ist vom Auftraggeber freizugeben.»

Tested: August 5, 2026

Bei here is locative. The acceptance happens at the manufacturer's works, which is what a factory acceptance clause says. Google made the manufacturer the party performing the test, a role the source never assigns, and no dictionary check catches it. DeepL avoided the error, but it attached the place to the test rather than to the acceptance, and the factory acceptance reading came from HardTrans. DeepL did step around the softer trap where Protokoll becomes “protocol,” which in English is the method, while the signed result is a report. Acceptance is not the only defensible rendering of Abnahme, and the objection to Google is the role, not the noun.

The reserve passage went the same way. On spannungsfrei zu schalten, DeepL wrote “de-energized” and Google “switched off.” A machine that is switched off is not safe to open, which is why EN 50110-1 opens its five safety rules with disconnecting.

An AI translator for large documents starts by reading the job

On a 200-page set, memory matters more than sentence quality. We make a first pass over every file you upload, write a project brief and a glossary out of it, and hold both while we translate each file. That is why “heat” stays “heat” on page 12 and on page 480. Upload your own glossary and we use your terms across the job; what it does not cover goes into the glossary we build, and you download that as CSV at the end.

An earlier run made the same point against Google. On a German pump spec, Google wrote “an output of 75 kW” in one sentence and “The performance of the drive” in the next, and turned the normative shall into a literal must. HardTrans wrote “a power output of 75 kW” and “The power output of the drive,” and kept “shall.” With your glossary applied it also locks the steel grade 1.4541 as X6CrNiTi18-10 (AISI 321). That pair sits with legal, medical and scientific ones on our side-by-side examples page; the mechanics are in why machine translation breaks on technical documents.

Two limits. No translation memory, no CAT infrastructure, and no human editor in the loop.

Is ChatGPT a DeepL alternative for documents?

For documents, no, and the reason is structural. ChatGPT returns text in a chat window. It does not hand you back a DOCX, XLIFF or PDF with the layout preserved, and it has no file pipeline for a multi-file set, so a 40-file job turns into 40 manual paste-ins.

Search “chatgpt translate pdf” and you get copy-paste workarounds and screenshots, never your file back with its layout intact. On the question people actually type, is ChatGPT good at translation, we have no number to give. We did not run these five passages through it, and we will not guess.

Translate a document without changing the format

You get back the file type you sent. DOCX, XLSX and PPTX keep their tables, numbering, lists and section structure, so you can translate a document without changing its format. For a PDF with a text layer we translate in place, so the PDF keeps its layout instead of coming back as a wall of paragraphs. XLIFF returns with inline tags intact, including .xliff, .sdlxliff (Trados), .mqxlz and .mqxliff (memoQ), .mxliff (Phrase), and Smartcat exports; the round trip is covered in our guide to translating XLIFF online. Scanned PDFs with no text layer are not supported.

Check it on your own document, first 10 pages free.

Pricing and turnaround

The figures come from 15 multi-file jobs of 100+ pages, March to July 2026, median 0.34 minutes per page. A 10-page document finishes in about 5 minutes.

When DeepL is all you need

Most of the time. A supplier email, a datasheet before a call, two or three pages that stay inside your team, a general engine is the right tool. Drift over three pages costs nothing.

It starts costing at scale and under signature. Twenty pages in, a reader who runs into three names for one assembly stops trusting the document. On a certificate or an acceptance clause, the one wrong term is the whole point of the sentence.

Common questions

Can HardTrans replace DeepL for CAT and XLIFF workflows?

For the pre-translation step, yes. Export XLIFF from Trados, memoQ, Phrase or Smartcat, translate it here, and import the result with its inline tags intact. It does not replace the CAT tool itself.

What does a page cost?

$1.50 per page of 250 words, covering the terminology pass over the whole job, translation in context, and formatting. Your first document up to 10 pages is free, no signup and no card.

Is the output good enough to send to a client?

As a draft, yes, and it cuts an editor's work substantially. As a final signed document, no. Have a specialist review anything that carries liability.

Upload the file you are actually stuck on. First 10 pages cost nothing.

Method. DeepL, one manual run in the DeepL web translator (German → English US), output pasted verbatim. Google Translate, public endpoint, three runs per passage, stable. HardTrans, production pipeline (PASS 1 + PASS 2, Claude Sonnet 4.6), single passages, no glossary. All runs August 5, 2026. DeepL plan prices checked August 5, 2026.