DeepL vs ChatGPT for Translation: Which to Use for Each Client Document Type

Neither wins on every job, so route by document type after a test on your own files. DeepL's file upload is designed to keep layout and apply glossaries, which suits contracts, manuals and slide decks. ChatGPT follows written instructions about audience and tone, which suits marketing copy. Score both on the same contract and the same marketing page before committing.

That is a starting hypothesis built from what each product is designed to do, not a test result, and no scores are given here: the outcome depends on your documents, your language pair and your reviewer. What follows is the test itself, the plan limits that decide which DeepL tier you need, the data questions to settle first, and a routing rule you can put on one page.

Follow me on Instagram@sagnikteaches

Two documents, ten planted traps and one reviewer

A contract and a marketing page pull in opposite directions. Contracts reward literal precision: defined terms, amounts, obligations. Marketing rewards adaptation: tone, rhythm, idiom. The same tool can win one and lose the other, so the test needs both. Pick one language pair where you have a bilingual reviewer who did not take part in producing either translation. English to Spanish is used as the example here.

Connect on LinkedInSagnik Bhattacharya

Build the two documents yourself instead of borrowing a client's. Document A is a services agreement of about 2,400 words and six pages, with numbered clauses, defined terms, a fee table, dates and amounts. Document B is a service page of about 600 words with a headline, three subheads, a bulleted list, a call to action and a couple of idioms. Replace real names with placeholders. If you do want to use genuine client material, check what the client contract allows before you upload anything to either tool.

Subscribe on YouTube@sagnikteaches

Then plant traps, so that errors have somewhere to show up. Ten that work:

• A defined term in capitals used a dozen times (for example "Service Credit"), to test consistency.
• One sentence with "shall", one with "may" and one with "will", to test whether obligation, permission and future fact keep their different strength.
• A number with separators, such as 1,250.00, and a percentage, such as 12.5%.
• A date written out in full, and a deadline expressed as "five business days".
• A cross-reference to "clause 7.2(b)".
• A negation with an exception: "shall not be liable unless".
• A table with merged cells, and a header or footer carrying the company name.
• A product name that must stay in English.
• A marketing idiom such as "hit the ground running".
• A slogan that depends on a pun and cannot be translated word for word.

Keep a private list of where each trap sits. The reviewer scores faster and more fairly when you can point them to the ten places that matter, after they have first read each output cold.

Set both tools up the same way before pressing translate

Most unfair comparisons come from unequal setup, not from the tools. Give each tool the same glossary, the same formality instruction and the same source file. Prepare a glossary of about 20 terms for the contract and ten for the marketing page, written as source term and agreed target term, and keep it in a spreadsheet so it can go to both tools unchanged.

In DeepL, upload the Word file rather than pasting text, attach the glossary, and set the formal or informal tone if it is offered. DeepL's plan table has a row for it, and the option does not appear for every target language, so check that it shows for your plan and language pair. On Team and Business plans you can also add shared style rule lists. In ChatGPT, create a Project, put the glossary in a Project file, and use one fixed brief for every run. OpenAI's help centre says a Project holds up to 5 files on Free, 25 on Go and Plus and 40 on Pro and Business, and that Project instructions override your general custom instructions inside that Project. Keep the glossary in a file, where one update reaches every chat in the Project.

You are translating a services agreement from English into Spanish
for a business client. Audience: the client's finance and legal teams.
Register: formal ("usted").

Rules
1. Translate everything, including headings, table cells, footnotes and
   cross-references. Do not summarise or shorten.
2. Keep numbers, dates, currency symbols and clause numbers as in the
   source. Change only decimal and thousands separators to the target
   convention.
3. Use the attached glossary for every listed term. If a term appears in
   another grammatical form, keep the glossary wording and adapt endings.
4. "shall" is an obligation, "may" is permission, "will" states a future
   fact. Keep those differences in the translation.
5. Do not translate: [product names, one per line].
6. After the translation, list (a) every place you were unsure,
   (b) any term not in the glossary that you had to choose, and
   (c) any sentence whose structure you changed.

Point 6 is the most useful line in the brief. A typical flagged item reads like this (invented for illustration): "Clause 9.2: 'reasonable endeavours' has no exact equivalent. I used a phrase meaning 'reasonable efforts'. Confirm with the client's lawyer." A query list turns the tool's uncertainty into a to-do list for the reviewer, and it shows quickly whether the model understood the document or merely converted it.

Run ChatGPT twice on each document. Its output can differ between runs, and a one-off lucky result is not a fair basis for choosing it. Note which run you carry forward and why. DeepL is not steered by a brief in the same way, so one run per document is a reasonable start, though a second pass after editing the glossary shows whether your entries took effect.

A scoring sheet a bilingual reviewer can fill in blind

Label the outputs A and B in a neutral file name and shuffle which tool is which for each document. The reviewer scores each criterion from 0 (unusable) to 5 (send as is), and marks errors on the text itself. Use two severities and keep them simple: a major error changes the meaning, an amount, a date or an obligation; a minor error is stylistic or a small terminology slip. Count both, but report majors on their own line, because one major in a contract outweighs a page of minors.

Weights differ by document type. Contracts spend most of their weight on meaning, terms and numbers. Marketing spends it on fluency and tone.

The weighted score is the sum of weight times score, divided by five, which gives a result out of 100. Record the minutes it took from opening the file to a version you would send to a client, because tool differences often show up in cleanup time rather than in raw quality.

Turning the reviewer's marks into one number per document

Give the reviewer one blank sheet per output, labelled only Output 1 and Output 2, and keep the key to which tool produced which until the marking is finished. The arithmetic is the same for both document types; only the weights change. Here is the sum for one contract output, with marks chosen to show the method, not taken from either tool:

Criterion            Mark   Weight   Points
Meaning                4  x   30   =  120
Terminology            4  x   25   =  100
Numbers and dates      5  x   20   =  100
Formatting fidelity    3  x   10   =   30
Fluency and tone       4  x   10   =   40
Consistency            4  x    5   =   20
Total                                 410
Weighted score  410 / 5 = 82 out of 100

Read the error log before the total. Two outputs within about five points of each other are a tie on a sample this small, and what separates them then is the majors line: one dropped "unless", one wrong amount or one "shall" turned into "may" in a contract outweighs a page of stylistic quibbles. If an output has a major error that the reviewer would not have caught without the trap list, note it separately, because that is the error that would have reached the client.

Add the clock. Time each run from opening the file to a version you would send to a client, step by step, and keep the log blank until every run is done so that one tool's times do not colour how you work with the other.

The minutes turn into money quickly. At $35 an hour for the reviewer, every ten minutes saved on a document is worth about $5.83, and across 40 documents a month that is about $233. That is why an output with slightly lower marks but a much shorter cleanup can still be the right choice for a job type, and why the routing table further down asks for both numbers. If one tool wins the contract and the other wins the marketing page, route by document type. If one leads on both, standardise on it and keep the other for the occasional exception.

What the planted traps look like when they fail

Reviewers mark faster and more consistently when they have seen what a failure on each trap looks like. The examples below are invented to show the error types for English to Spanish; they are not output captured from DeepL or ChatGPT.

Give the reviewer this table before the marking starts, together with the private list of where each trap sits. A trap that both outputs pass tells you less than one they split on, so when the results come back, look first at the traps where the two outputs differ.

Glossary control: run a term audit instead of trusting the setting

A glossary setting tells you what you asked for, not what you got. The audit takes ten minutes. Search the English source for each glossary term and count the occurrences, then search each output for the agreed target term and count again. The two counts should match, and every gap is a violation to fix or a reason to distrust the tool for that job.

The source counts come from your own planted contract; the two output columns stay empty until the runs are done, and each total is the share of the 31 uses that came out right. A result such as 26 of 31 is 84%, and every miss is a place the reviewer has to catch by hand. The audit exposes a difference the plan page hides. DeepL's plan table lists the number of glossary entries per language pair: 5 on Free and Individual, 1,000 on Team, unlimited under a fair usage policy on Business. A 20-term contract glossary does not fit on Free or Individual at all. ChatGPT has no entry cap in the same sense, but it only follows the glossary as an instruction, and instructions slip in long documents. If you want a fuller method for building the term list itself, how to build a product glossary AI must use in every draft is the place to start.

Plan limits that decide which DeepL tier you need

DeepL has restructured its plans: the current ones are Free, Individual, Team, Business and Enterprise, and older articles that list other tier names are out of date. For an agency, the character allowance matters less than the file allowance, because the number of file translations runs out first.

The prices are the USD prices DeepL's pricing page showed in early November 2026, per user a month on annual billing. DeepL sets prices by country and shows some regions a local currency, so confirm yours there. The limits come from the plan table on the same page. Team adds five shared style rule lists, team administration and bulk file translation; Business adds translation memory, 20 style rule lists and the Write Pro writing tool, which Individual and Team can only buy as an add-on (listed at $7.49 per user a month, billed annually).

Here is how the limits bite. Take a three-person agency where each translator handles eight contracts and six marketing pages a month. That is 14 files and roughly 137,000 characters per person, using about six characters per word. Characters are not the constraint at all: Individual's three files a month fails immediately, and Team's 20 files fits comfortably. If the same translators each handled 30 files a month, Team would run out and the choice would move to Business or to the DeepL API, whose plans I would only consider if you connect to a CAT tool.

For the API, DeepL lists a free Developer tier with a one-time credit of 1 million characters, and a Growth tier at $26 a month billed annually, with 12 million characters a year included and $27.50 for each extra million. Growth lists CAT tool integration, 2,000 glossaries and no data training, and neither API tier includes the web translator. ChatGPT is simpler: OpenAI's help centre lists Plus at $20 a month per person, and its business pricing page lists ChatGPT Business Standard seats at $20 a user a month billed annually or $25 billed monthly, for teams starting at two users.

Put the subscription gap in context. For three translators, DeepL Team is $86.22 a month and ChatGPT Plus is $60. The difference is about $26. Now price review time: 40 documents a month at 1.5 hours each and $35 an hour is $2,100 of reviewer time. A 10% change in review time is worth $210, which is about eight times the subscription gap. Tool cost rarely decides this; the review hours do.

Formatting: what to check in a returned file

DeepL says its document translation handles 13 file types without conversion, including Word, PowerPoint, Excel, PDF, HTML, SRT and TXT, keeps the original formatting, and uses OCR (optical character recognition) to read scanned PDFs and images while preserving their layout. Treat that as a claim to test on your two documents. Open each returned file next to the source and check six things: table borders and merged cells, headers and footers, numbered clauses and cross-references, fonts, text that has grown longer than the space it sits in (common on slides), and anything inside text boxes.

ChatGPT is a different workflow. If you paste text, the output is text and you rebuild the layout, so count that time. If you upload a Word file, whether you get a finished file back may depend on your plan and how you ask, so test it. For image-only scanned PDFs, run an OCR step before ChatGPT sees them, or test a sample first, and give DeepL's scanned-PDF handling the same sample before relying on either. For a deeper look at the ChatGPT side of file work, see translating a PDF with ChatGPT.

Data handling: settle this before you upload anything

Client contracts often limit which third parties may see their documents. Check yours first, then compare what each vendor says in its own words. DeepL lists "No data training" on its paid plans, and its pricing FAQ says that for Pro subscribers it stores original content and results only as long as technically needed, deletes them afterwards unless otherwise specified, and never uses them to train its AI models. The pages I read do not describe the free version's handling in the same terms, so I would not put client material in it. OpenAI's business pricing page says ChatGPT Business content is not used to train its models, and its consumer pricing page shows an opt-out for training on Free, Go, Plus and Pro.

One more ChatGPT detail matters if a client contract says where data must be stored: OpenAI's business pricing page lists data residency for Enterprise, not for Business. Settle that before you choose a plan, and ask DeepL the same question in writing. None of this replaces a read of the current terms, and the vendor pages change, so treat what I have written as a checklist, not as a legal opinion. The wider safety questions are covered in whether it is safe to put client documents into AI translation tools.

Routing jobs once the scores are in

Write the routing rule as a table that a new team member can follow without asking you.

For the post-edit step itself, a translator's machine-translation post-editing workflow sets out the passes. If you want a more generic version of this test that works for any tool, an accuracy test on your own documents covers it.

Mistakes that spoil the comparison

Four slip-ups are common enough to check for. The first is unequal inputs: you give ChatGPT the glossary but forget to attach it in DeepL, then conclude DeepL ignores terminology. The second is a reviewer who knows which output is which, which quietly biases the scores towards whichever tool they expected to win. The third is judging only the first page, where both tools look best, and missing the drift in clause 14. The fourth is ignoring cleanup time, so that a tool with slightly better text but a broken layout appears to win.

A fifth deserves its own line: do not test with text that is already public. A famous contract template or a well-known web page may be familiar to the model from its training, which flatters its result. A private document that you wrote or adapted tells you what happens on new material.

A seven-step version for this week

1. Write the two documents and plant the ten traps. Strip any real names.
2. Build a 20-term contract glossary and a ten-term marketing glossary.
3. Find a bilingual reviewer who did not make either translation.
4. Set up DeepL (file upload, glossary, formality) and a ChatGPT Project with the same glossary and the brief above.
5. Run both tools, ChatGPT twice, and log the minutes for each step.
6. Have the reviewer score blind, then run the term audit yourself.
7. Fill in the routing table and save it with the scorecards. Repeat the test in six months, because both vendors change their plans and models often.

Questions agencies ask before they pick a tool

Can I use the free versions for client work?

Only after reading the terms against your client contract. DeepL Free is capped at 50,000 characters a month, one file translation and five glossary entries per language pair, which rules it out for agency volumes anyway. Personal ChatGPT plans offer an opt-out from model training. Paid DeepL plans list no data training, and OpenAI says ChatGPT Business content is not used for training.

Will ChatGPT keep a document's formatting?

It may, depending on the plan and how you ask, so test it rather than assume. If you paste text, rebuilding headings, tables and numbering is your job, and that time belongs in the comparison. DeepL's file upload says it preserves formatting and layout for Word, PowerPoint, Excel and PDF files, but check tables, headers and slide overflow in the returned file.

How many documents make a fair test?

Two documents are a pilot: one contract and one marketing page. For a decision that affects your whole workflow, use five or six, a mix of both types, with the same bilingual reviewer scoring blind. Treat differences of a few points as noise and look at the severity of errors instead, because one wrong amount matters more than ten stylistic quibbles.

Is DeepL's glossary stricter than a ChatGPT prompt?

DeepL applies a glossary as part of the translation step, while ChatGPT follows the glossary you paste into the prompt or a Project file. Neither is guaranteed to obey in every sentence. The answer for your language pair comes from a term audit: count how often each agreed term appears correctly in each output, and compare the two percentages.

Further reads

• How Translators Use AI to Keep Terminology Consistent — Keeping terminology consistent across a project and across translators.
• Translation Memory and AI: How to Combine Them in a Small Agency — How translation memory and AI work together in a small agency.
• MTPE Rates per Word: How Translators and Agencies Price Post-Editing — Pricing post-editing once a machine draft is the starting point.
• Does Your Translation Business Need an AI Usage Policy? — Writing the AI usage policy your clients will ask to see.
• AI Legal Translation: What a Certified Translator Can Use and Disclose — What to use and disclose on certified legal work.
• memoQ vs Trados for Small Translation Agencies: AI Features and Cost — Where AI features sit inside the main CAT tools.
• Tour Descriptions in Several Languages: An AI Workflow — A repeatable workflow for tour descriptions in several languages: one master, a facts table, AI rewrites, three translation checks and channel versions.
• AI-Powered Translation Management System: What a Small Agency Needs — Which translation management system suits a five-person agency: published prices, a ranked shortlist, an exit test and a two-week trial plan.
• AI Subtitle Translation for Agencies: Quality Checks Before Delivery — A cue-by-cue workflow for AI subtitle translation, with timing, line-length and meaning checks, a QC sample and a brief for the native reviewer.
• ISO 18587 Certification: Is It Worth It for a Small Translation Agency? — What ISO 18587 asks of post-editors and process, how it relates to ISO 17100, the audit steps, and a break-even test for a small agency.
• Is AI Medical Translation Safe for a Practice? Rules and Workflow — AI can translate reminders and directions with review, but consent, clinical instructions and live consultations need a qualified interpreter or translator.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: DeepL pricing page (plan table and FAQ), document translation page and API pricing page; ChatGPT pricing page, OpenAI business pricing page, and OpenAI help centre articles on ChatGPT Plus and Projects. Prices read in early November 2026; DeepL sets prices by country.