A DeepL alternative for documents

Author: Yuri Vodostoy — Technical translator and founder of Insight Translation · Updated 2026-08-26

HardTrans reads every file in a job before it translates a word, builds a project brief and a glossary from it, and holds both across every file. DeepL translates what you hand it. On a German regulation run through all three engines on August 25, 2026, Google Translate needed 13 different English forms for six recurring terms, DeepL 8; HardTrans used exactly one form per term, six for six. On five engineering passages three weeks earlier, DeepL slipped where a standard fixes the term; Google got none of the five clean. $1.50 per 250-word page, first 10 pages free.

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 a 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, and three of them are on this page.

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.

The English text of Directive 2009/104/EC splits it the same way: Article 2(b) reads Instandsetzung, Umbau, Instandhaltung und Wartung in German and “repair, modification, maintenance and servicing” in English. Checked on EUR-Lex, August 25, 2026.

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.

The long-document run three weeks later says it again as a number. On one eight-paragraph regulation, DeepL held six recurring terms to 8 English forms and Google Translate to 13.

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.

Technical documents: the term has to hold on page 12 and on page 480

One sentence at a time, all three engines are good. They come apart over length. On August 25, 2026 we put § 14 of the German BetrSichV, the rules on inspecting work equipment, through DeepL, Google Translate and HardTrans: 4,291 characters, eight paragraphs, six terms that repeat between 3 and 32 times. We did not count quality. We counted one thing: how many different English forms one German term came back in.

Source, paragraph (2): «Arbeitsmittel, die Schäden verursachenden Einflüssen ausgesetzt sind, die zu Gefährdungen der Beschäftigten führen können, hat der Arbeitgeber wiederkehrend von einer zur Prüfung befähigten Person prüfen zu lassen. Die Prüfung muss entsprechend den nach § 3 Absatz 6 ermittelten Fristen stattfinden. Ergibt die Prüfung, dass ein Arbeitsmittel nicht bis zu der nach § 3 Absatz 6 ermittelten nächsten wiederkehrenden Prüfung sicher betrieben werden kann, ist die Prüffrist neu festzulegen.»

Tested: August 25, 2026

The highlighting marks the two terms this paragraph tests, not mistakes. Wiederkehrend occurs twice in three sentences, and DeepL rendered it once as “at regular intervals” and once as “periodic”; over the whole text it used periodic six times and regular once. It also split zur Prüfung befähigte Person into “qualified inspector” four times and “person qualified to conduct inspections” three times. Google split that term three ways, ending with “person authorized to carry out the examination.” HardTrans used “competent person” seven times out of seven.

German → English, tested August 25, 2026. A form is the lexical choice for the head term; derivatives such as Prüffrist and Prüfumfang are counted separately and are not in this table.

DeepL is the more consistent of the two general engines, 8 forms against 13, and its drift is narrow: one modifier, one compound term. Google's is on the head term of the document. In paragraph (5) alone die Prüfung comes back as examinations, exam, examination, test and inspection: “The deadline for the next recurring exam begins with the due date of the last exam.”

The text was cut up because DeepL's free web window takes 1,500 characters and the source is 4,291. Both engines got the same four chunks, 1,180, 736, 1,370 and 999 characters, pasted one after another in a single tab. Google's window would have taken the whole text at 5,000, but the narrower engine sets the track for both. HardTrans got the file whole, because that is what it takes: a file, not a paste.

Which leaves the obvious objection, so we ran ours handicapped too. The same four chunks as four separate jobs, each with its own first pass, nothing carried between them. Six forms became eight, level with DeepL, and the term that broke was wiederkehrend, split into recurring, periodically and periodic exactly the way Google split it. The margin in the table is not a better engine, it is having read the whole document before translating any of it, which is the part described in why machine translation breaks on technical documents.

Two things that number is not. It is not our repeat handling: we collapse exact duplicates before translating, and this text has none, so nothing was collapsed. And it is not a count of errors, because English tolerates variation and three forms are not three mistakes. On one of the six terms there is a right answer. BetrSichV implements Directive 2009/104/EC, whose English text calls the person doing the inspection a competent person (Article 5). That is the form HardTrans used; DeepL and Google each coined their own.

The test has limits of its own. This is a regulation, so the register is legal rather than shop-floor, and a maintenance manual would put different terms under pressure. DeepL's output is also not reproducible word for word: two runs of the same four chunks on the same day differed in two places, “Exceptional events” against “Extraordinary events” and “carried out on time” against “in a timely manner.” The file we kept is the second run.

The source is § 14 BetrSichV on the German federal legal portal, taken verbatim with HTML sublists flattened to plain lines. It has no official English version, so it is less likely to sit in the parallel corpora these engines learned from. Anyone can rerun it.

Your terminology, where the standard leaves a choice

Some terms have no single correct English form. The standard allows more than one, and the house term decides, which is what a glossary is for. DeepL has one too, so we gave both engines the same eight-entry German-English glossary and the same four lines of a pump specification.

Source: «Die Pumpe ist für eine Leistung von 75 kW auszulegen. / Die Leistung des Antriebs ist im Abnahmeprotokoll zu dokumentieren. / Das Gehäuse besteht aus dem Werkstoff 1.4541. / Alle Dichtungen sind vor der Abnahme zu prüfen.»

Tested: August 25, 2026. The same glossary on both sides.

Start with the part that doesn't flatter us. With the glossary loaded, DeepL produced almost exactly our output, term for term. On four lines a glossary works on both sides, and a page that told you otherwise would be lying to you. Google Translate has no glossary control in its free web interface at all, checked the same day; terminology lives in the paid Cloud Translation Advanced, a different product we did not run. It reached “power output” on its own anyway.

Our own weak spot is the third panel. Without a glossary that second line was unstable: three runs of the same sentence returned “The power of the drive” twice and “The power output of the drive” once. With the glossary, all three runs came back identical. The highlighting here marks the house term against a different but defensible choice, not right against wrong. Housing is ordinary English for Gehäuse; pump documents say casing.

What decides which tool goes on a job is the quota. DeepL's free web plan took five of our eight entries on August 25, 2026 and refused three, on quota rather than content. The glossary control itself is not behind Pro, but the creation wizard accepts three term pairs, you get one glossary per account, and five entries per language pair. Generating a glossary from past translations is a Pro feature and we did not use it. We say nothing about DeepL Pro or its document mode, because we did not run them.

The cap cost DeepL nothing here, and it is worth saying why: the three refused entries, Antrieb, Abnahme and Abnahmeprotokoll, were ones it already rendered our way by default. On a real job it is five terms out of however many the document needs. To keep the comparison level we also ran our side with only those five entries, and the output was identical to the eight-entry run.

Our glossary takes CSV, TSV, TXT or a Smartcat XLSX, several files at once, and applies across every file in the job rather than to one paste. Before you pay, the estimate shows the term count, any conflict where one source term has two targets, and a warning when your glossary barely appears in the document. When the job finishes you download two CSVs: the terms that were used, and the glossary the first pass built where yours was silent. The same steel grade sits with legal, medical and scientific pairs on our side-by-side examples page.

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, 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. On the question people actually type, is ChatGPT good at translation, we have no number to give. We did not run these passages through it, and we will not guess.

What we do to a PDF, and what we don't

We take a PDF that has a text layer and give you back a PDF, translated in place: the source text is removed and the translation is written into the same box. A scan holds a picture of characters rather than characters, and that one fact decides what any engine can do with it, ours included.

What in-place translation costs is worth knowing before you upload. Where we can't resolve the original font we substitute DejaVu Sans, so the typeface changes. Where the English runs longer than the German the type is squeezed, down to half the original size, so a tight table or a dense drawing will not come back looking like the original. The repeat discount doesn't apply to PDFs either, so the same text costs more as a PDF than as a DOCX.

A scan with no text layer is not “not recommended,” it is refused: the word count comes out at zero and the upload returns an error. We don't run OCR. Images inside a PDF are left exactly as they are, any text in them included, and the file is flagged before you pay.

So if the layout is the point, PDF is the wrong input. Send the DOCX or the XLIFF the PDF was made from. DeepL's plans meter document translations — one a month free, then 3, 20 or 100 depending on the plan, checked August 26, 2026; we have not run their document mode and make no claim about what it does to a PDF.

What you get back

The file type you sent. DOCX, XLSX and PPTX keep their tables, numbering, lists and section structure, and XLIFF comes back with inline tags intact, from .sdlxliff and .mqxliff to Smartcat exports; the round trip is in our guide to translating XLIFF online. What survives format by format, and how the four kinds of tool compare on it, is set out in what makes an AI translator good for documents.

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

Pricing and turnaround

$1.50 per page of 250 words, and the first 10 pages cost nothing.

The speed 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. Text that repeats inside a file is billed at 25%; that discount is per file, not across the job, and it does not apply to PDFs.

As a draft, the output saves an editor real time. As the final signed document it isn't one. There is no human in our loop, so anything carrying liability needs a specialist to read it before it goes out.

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.

Can I use my own glossary, and does DeepL have one?

Ours takes CSV, TSV, TXT or a Smartcat XLSX, several files at once, applied to every file in the job. Conflicts are flagged before you pay, and the terms used and the glossary we built download as CSV. DeepL's free web plan took five of our eight entries on August 25, 2026: one glossary, five entries per language pair.

Can DeepL translate a PDF, and can you?

The text layer decides. We accept a PDF that has one and return a PDF translated in place, with the font substituted where we can't match it and the type squeezed where the English runs long. A scan with no text layer is refused at upload, and we don't run OCR. We have not run DeepL's document mode, so we make no claim about it.

What happens to my files?

Files are deleted from our servers automatically after 7 days. Your text is never used to train models — neither by us nor by Anthropic, the model provider. Anthropic retains API data for a limited period under its policy and does not train on it.

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

Method. August 5, 2026: 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, single passages, no glossary. August 25, 2026, consistency run: source § 14 BetrSichV, 4,291 characters; DeepL and Google Translate in their free web windows, four chunks of 1,180/736/1,370/999 characters set by DeepL's 1,500-character limit, one tab, no reload, output read verbatim from the result pane and cross-checked against each engine's own copy button; HardTrans on the whole file in one job, plus a control run of the same four chunks as four independent jobs. August 25, 2026, glossary run: the same eight-entry German-English glossary offered to both engines, DeepL's free plan accepting five of them, three runs per state on our side. DeepL plan prices checked August 5, 2026.