How Many Words Can a Translator Translate Per Day? Realistic Numbers for 2026

Ask ten linguists how many words a translator can do in a day and eight will say 2,000 to 2,500. That range has been repeated so long that almost nobody checks it against their own logbook. We have watched the same translator, same language pair, same client, produce 4,100 words on a Tuesday and 900 on a Wednesday. The translator words per day figure worth planning around is not the one you can reach on a good day. It is the one you can repeat on an ordinary one.

This is not idle curiosity. Your daily output decides which deadlines you can accept, what your real hourly rate is, and whether a rush job is profitable or a quiet loss. Below are the numbers we see in practice, split by domain and by workflow, plus a way to find your own instead of borrowing someone else's.

Where the 2,000 words per day benchmark came from

The figure predates almost everything in a modern workflow. It settled into the industry when translators worked from paper source, typed the target, and had no translation memory, no termbase lookup and no concordance search. Two thousand words was a fair day of typing plus dictionary work, and it happened to divide neatly into the per-word pricing model agencies were building at the same time.

That second part is why it survived. Project managers need a capacity number to schedule against, and 2,000 a day gives clean arithmetic: a 10,000-word file takes a week, minus a day for review. Nobody derived it from a study of how fast people work across different content types. It was a scheduling convention that hardened into a belief.

The convention still earns its keep. When a client asks how long something will take and you have no history with that content, quoting from a conservative baseline protects you.

The trouble starts when the number gets used in the other direction, as a floor rather than an estimate. We regularly hear from freelancers who think they are slow because a dense pharmacovigilance report took two days at 1,100 words a day. But 1,100 was the correct speed for that document. Anything faster would have meant skipping terminology verification, which is not speed, it is a deferred cost.

Treat 2,000 as a historical average with enormous variance underneath it.

What actually sets your daily word count

Most of your output is decided before you translate a sentence. The biggest variable is how many real decisions the text forces per hundred words. A software release note with a client glossary attached might ask three genuine questions per paragraph. A supply contract with defined terms, cross-references and a governing-law section asks one every second line, and each one has consequences.

File condition comes next, and it is badly underrated. A clean DOCX with proper heading styles behaves. A DOCX that started life as a scan and went through OCR arrives with broken paragraph breaks, hyphenated line endings, text boxes where tables should be, and stray characters that look like terminology until you check them. We have seen an 8,000-word manual take one translator two days in one format and four in another. Identical words.

Reference material changes the picture again. With a client glossary, a previous version of the document and a style guide, you translate. Without them you research, then translate, then go back and harmonise the terms you guessed at in the first thousand words.

Source quality is its own tax. Text written by a non-native author, or assembled by a committee, makes you reconstruct intent before you can render it. None of that reconstruction shows up in the word count.

Subject familiarity compresses everything above. The tenth annual report for the same client runs at roughly double the speed of the first, and the gain comes almost entirely from not re-deriving the terminology and the register.

Realistic translator words per day by domain

These are ranges we see from working professionals for first-pass human translation, review time excluded, on a normal day rather than a deadline sprint.

Technical documentation and manuals sit highest, usually 2,500 to 3,500. Repetitive structure, controlled vocabulary and a settled glossary do most of the work. Machinery documentation with numbered procedures can push past 4,000 once the terminology is locked. That ceiling drops fast when the manual carries diagram callouts, unit conversions or safety warnings that need regulatory phrasing.

General business correspondence and reports land in the classic middle, 2,000 to 3,000. Register matters, research rarely does.

Legal work runs 1,200 to 2,000. Contracts, judgments and regulatory filings slow down because precision is not optional and because one defined term propagates through thirty pages. A translator we spoke to who works mainly on share purchase agreements counts a good day as 1,500 and treats anything above 2,000 as a warning that she has stopped checking cross-references.

Medical and clinical content is similar or slower, roughly 1,000 to 2,000. Study protocols, informed consent forms and device instructions carry terminology load and real consequence. Patient-facing material is slower again, because you are translating and adjusting reading level at the same time.

Marketing and transcreation break the model entirely, often 500 to 1,200. Forty words of headline copy that actually lands in the target market can take an afternoon, and counting them tells you nothing.

Two caveats worth stating plainly. These describe sustained output, not personal bests. And they assume you are not simultaneously managing the project, chasing reference files and answering client email out of the same hours.

One more thing the domain ranges hide: the first file for a new client in any of these categories runs slower than the range suggests, sometimes by half. You are learning their house terminology, their formatting habits and how literal they want you to be. Freelancers who quote a new client at their established speed for that domain are usually quoting for their third project, not their first. Add a day to the first one and stop being surprised by it.

How post-editing changes the arithmetic

Post-editing moves the numbers up, and by much less predictable amounts than vendors suggest. In our experience the honest range for full post-editing of good AI output is somewhere between 4,000 and 7,000 words a day, with light post-editing higher. Published MTPE productivity research spans a very wide band, and the width is itself the finding: results depend on raw output quality, domain, language pair, and how disciplined the post-editor is about repairing versus rewriting.

Two things distort the averages. First, raw quality is not uniform inside a single file. A manual might arrive with 70% of segments needing a light touch and 30% needing a full rewrite, and those rewrites eat the time the easy segments saved. Averaging across the file hides exactly the part that hurt.

Second, fatigue. Post-editing is a different cognitive task from translating. You are reading against a proposal rather than generating from scratch, which is faster and considerably more monotonous, and error detection drops off after a few hours. Someone who clears 7,000 words in a day is usually reviewing the last 1,500 less carefully than the first. If you work at that volume, splitting the day and running a separate quality pass over the final stretch is worth the hour it costs.

There is a related trap in how these jobs get quoted. Because the raw output is already there and reads fluently, both the client and the post-editor tend to estimate from the fluency rather than from the accuracy. A file that looks 90% done can still hide a mistranslated liability clause or a unit that was silently converted. Catching that costs the same amount of attention it would have cost in a human-translated file, and it does not scale with the throughput number on your invoice.

Language pair matters more here than it does in human translation. Post-editing English into German or Spanish is a different job from post-editing English into Japanese or Finnish, where restructuring the sentence is routine rather than occasional. The productivity gain narrows sharply, and quoting a single MTPE day rate across all your pairs is how post-editors end up subsidising their hardest work with their easiest.

Why your best day is the wrong number to plan with

Almost every capacity mistake we see traces back to the same habit: quoting a deadline from the fastest day you can remember. Peak output is real. It is also not repeatable, because it happened on a day with no interruptions, clean source material and a subject you knew cold.

Do the arithmetic properly. A 12,000-word technical file at a peak of 4,000 words a day looks like three days. Now add what the word count never includes: reading the source before you start, building or checking terminology, a revision pass at roughly a third to a half of translation speed, a formatting check on the delivered file, and the email thread. Five working days is closer, and that is before anything goes wrong.

There is a second cost that surfaces a week later. Two consecutive peak days tend to produce a third day at around 60% of normal, and the errors from day two show up during the revision of day three. Sustained overproduction borrows from next week at a bad rate.

One planning rule has held up well for us: quote from your median day, never your maximum, and be careful with your mean. Say your logged days over a month run 1,800, 2,400, 900, 3,100, 2,200. The mean is 2,080 and the median is 2,200, which look similar enough. But the 900 day is the informative one. It tells you what happens when a file arrives badly formatted, and it will happen again. Build a buffer for it rather than filing it under bad luck.

How to measure your own baseline

Borrowed benchmarks are worth roughly what borrowed rates are worth. Twenty working days of your own data is enough to plan with, and the tracking does not need to be clever.

Log four fields per session: start time, stop time, words completed, content type. Add a fifth for file condition, even a crude one, along the lines of clean, workable or awkward. That column will explain more of your variance than any other.

Keep first-pass translation separate from revision. Combining them gives you a number you cannot use for either quoting or scheduling, because the ratio moves by domain. Legal work might run three to one, translation to revision. A marketing piece can approach one to one.

Count words per productive hour rather than per day, then multiply by the translation hours you actually have. Most full-time freelancers translate for four to six hours out of eight, with the rest going to email, quoting, invoicing and file admin. Someone producing 500 words an hour over five hours has a 2,500-word day, and knowing the hourly figure lets you quote a half-day job properly instead of guessing.

After twenty days, sort by content type and take the median in each bucket. Those medians are your quoting numbers. For a fuller method, we covered building a productivity baseline separately.

Re-measure once a year. Workflow changes, particularly the addition of AI-assisted steps, shift these numbers enough that a two-year-old baseline is quietly wrong.

What to do with the number once you have it

Start with the deadlines you offer. Take the median for that content type, subtract 20% for the file arriving in worse shape than promised, add revision at a third of translation time, and quote from that. You will lose the occasional job to someone promising faster delivery. You will also stop delivering at 2am.

Then look at capacity. Twenty working days at your median gives your monthly ceiling in words. Multiply by your rate and you have the realistic maximum you can earn from translation alone. Most freelancers find that number lower than they expected, and the useful conclusion is that speed is the weaker lever. Going from 2,200 to 2,600 words a day is hard, costs quality and buys you 18%. Moving from a mid-market rate to a specialist rate buys more and does not require you to type faster. We wrote about that trade-off in pricing your services confidently in 2026.

So here is the concrete step. Open a spreadsheet today with five columns: date, content type, file condition, translation minutes, words completed. Fill it in for twenty working days without changing how you work. Then calculate the median words per hour for each content type, and use those medians for every quote you send for the next quarter. Note the two or three jobs where the estimate was wrong, and why.

Most translators find their real number sits well below the one they have been quoting. The gap between the two is where the unpaid overtime has been living.