Post-Editing Productivity: Realistic Words-Per-Hour Benchmarks

You ask three post-editors how fast they work and get three completely different numbers. One says 1,500 words an hour. Another says 400. The third says “depends on the day.” All three are telling the truth - which is exactly what makes productivity benchmarks in MTPE so difficult to use correctly.

Here’s what the numbers actually mean, where they come from, and why the average figure you find in articles is probably both right and useless for your specific situation.

The baseline: what human translators actually produce¶

Before talking about post-editing speed, you need the baseline: how fast does translation from scratch actually go?

The industry standard is 250-400 words per hour for professional human translation, which translates to 2,000-2,500 words per day on a standard working day. This is the figure used by the EU Translation Centre, professional associations, and most agency workflows.

Two important caveats:

First, this 250-400 range already assumes a professional translator working in a familiar domain with a good CAT tool. Put that same translator in an unfamiliar domain or a language pair with limited resources, and output can drop to 150-200 words per hour. Experienced specialists in routine content can push 500 words per hour from scratch.

Second, “words per day” includes time for research, TM queries, terminology lookup, and QA - not just the typing. The 2,000-word-per-day figure is net productive output, not uninterrupted keystrokes.

This baseline matters because every MTPE productivity claim is only meaningful relative to it. “Post-editing is 3x faster” means nothing unless you know the starting point.

Light post-editing: the 1,000 words/hour ceiling¶

The most-cited benchmark for light post-editing is approximately 1,000 words per hour - roughly 2.5-4x the human translation baseline.

Where does this number come from? It comes from studies on structured content with good MT output. A Migros Bank study measured 700 words/hour on financial texts. A Slator analysis cites translator Enrico Antonio Mion reporting 1,000 words/hour for light post-editing of well-trained MT output. News articles (Incyta LSP case study) reached 1,500 words/hour for lightly edited content.

But these conditions - good MT, familiar domain, structured content - are the best-case scenario.

The EU Translation Centre uses 20 pages per day (~5,000-8,000 words at 250-400 words/page) as its standard for light MTPE. In practice, this means the ceiling for light PE is around 5,000 words per productive day in a well-organized workflow.

When light PE actually hits these numbers: - MT engine trained on the specific content domain (not generic models on specialized texts) - Language pair where MT quality is reliably high (European high-resource pairs) - Content is formulaic and repetitive (product descriptions, UI strings, release notes) - Post-editor is familiar with the engine’s error patterns - No ambiguous or contested terminology that requires research

Miss any of these conditions and throughput drops. Most real projects miss at least two.

Full post-editing: the math that doesn’t save as much as agencies promise¶

Full post-editing (FPE) - where the output must be indistinguishable from human translation - runs at 600-800 words per hour, or 2,000-3,000 words per day.

The EU Translation Centre standard for full MTPE is 15 pages per day (~3,000-5,600 words depending on page density). Crowdin’s analysis pegs FPE throughput at 600-800 words/hour - roughly 2-3x the human translation baseline when MT quality is good.

Here’s where the math gets uncomfortable: at 2,000-3,000 words per day, an FPE editor isn’t dramatically faster than a human translator. They’re doing more work per word (reading the MT output, evaluating it, deciding what to change, then making the change - versus just translating directly). For many content types, this cognitive overhead partially cancels out the benefit of having a draft.

A GTS Translation survey of 212 freelancers in 2025 captures this reality directly:

“Around 50% of respondents do not offer discounts for MTPE work, arguing that post-editing can take as much time as traditional translation.”

And 66% of respondents said MT output requires significant edits, with 21.74% saying it requires extensive rework - effectively retranslation at a lower rate.

Full post-editing is genuinely faster than translation from scratch when MT quality is good and the content is structured. It’s not meaningfully faster when either condition fails. Agencies that sell FPE as “human quality at half the time” are often working from best-case benchmarks, not average-case reality.

The benchmark table¶

Source: SwissGlobal 2026 analysis, Crowdin guide, Weglot analysis.

The most important column is the right one. When MT quality is poor, you don’t just lose the speed advantage - you can end up slower than translation from scratch because the cognitive effort of evaluating and discarding bad output costs time you’d never have spent with a blank page.

Language pair is the biggest variable you can’t ignore¶

The overall average benchmark - “MTPE is 66% faster than human translation” - masks enormous variation by language pair. SwissGlobal’s 2026 analysis of 90 million words across 879 linguists is the most detailed public data on this:

For EN→FR, light PE at 1,000+ words/hour is achievable and the economics of MTPE are compelling. For EN→SV, the same post-editing workflow is slower than just translating - the MT output quality doesn’t justify the cognitive overhead of evaluating it.

A 2025 Finnish study on GenAI post-editing found an average speed gain of 14% over human translation from scratch - with individual variation from -2% to +102% across post-editors working on the same content. That’s not noise. That’s a signal that individual editor skill and engine familiarity matter as much as the language pair itself.

Practical implication: never set a productivity expectation for MTPE without first running a pilot in the specific language pair. Using EN→FR benchmarks to set expectations for EN→JP will consistently produce project failures.

Content type effects on throughput¶

Content domain affects productivity almost as much as language pair. The reason is simple: MT accuracy varies dramatically by domain and text type.

From published LSP case studies and research:

High-throughput content types (light PE at 1,000-1,500 words/hour): - Software UI strings and application help text - E-commerce product descriptions with standardized formats - News articles on predictable topics (sports scores, weather, press releases) - Technical documentation with controlled vocabulary

Medium-throughput (600-900 words/hour): - General technical manuals - B2B marketing materials - Financial reports with standardized structure

Low-throughput (250-600 words/hour, approaching human translation speed): - Legal contracts with complex sentence structures - Medical documents requiring clinical accuracy verification - Literary or highly stylized content - Highly localized content where MT doesn’t capture cultural register

The Migros Bank study on financial texts is illustrative: 700 words/hour for structured financial translation - lower than general tech content, because MT models trained on general text tend to handle financial terminology less reliably and terminology errors require research to fix.

Literary translation shows the widest range in the data: 402-1,140 words/hour in one study. That extreme variance tells you productivity in complex content depends heavily on the specific text and post-editor, not just the workflow.

The experience factor: why the same workflow produces different outputs¶

Post-editor experience with a specific MT engine matters significantly - and not in an obvious way.

According to data from Slator’s analysis, post-editors typically reach their maximum productivity after approximately three months of regular work with the same MT engine. Before that, they’re discovering error patterns as they go - which means reading more carefully, catching more surprises, taking longer.

After three months, an experienced post-editor has an internalized map of where the engine fails: - Which terminology it consistently mistranslates - Which sentence structures it garbles - Which content types trigger poor output - Where it gets close enough to accept with minor edits vs. where to expect major rewrites

This front-loaded experience curve has a direct business implication: the first two months of a new MTPE workflow are always less productive than the benchmarks suggest. Factor that into project timelines and rate negotiations for new engines or new language pairs.

General translation experience helps, but it’s not the same thing. A translator with 15 years of EN→DE experience who’s new to a specific MT engine will still be slower than a 3-year post-editor who knows that engine’s DE output patterns. The skill is partially transferable, but engine familiarity is its own domain.

When post-editing doesn’t save time: the scenarios to know upfront¶

There are project configurations where MTPE provides zero productivity benefit - and a few where it actively costs time. These aren’t edge cases; they’re common enough that every agency should know them.

Scenario 1: Domain mismatch. You use a general MT engine on highly specialized content (patent claims, niche legal terminology, medical device documentation). MT accuracy is low enough that post-editors spend more time identifying problems than they would have spent translating. Light PE becomes effective FPE, and effective FPE approaches retranslation.

Scenario 2: High-resource pair pushed through a low-resource-trained engine. A GPT-era general model on EN→FI or EN→HU can produce output that’s grammatically almost correct but systematically wrong in ways that take more time to evaluate than to translate directly.

Scenario 3: First project with an unknown engine. Before an editor builds familiarity with an engine’s error patterns, productivity is at its lowest. Combined with a new language pair or domain, you can easily end up below human translation throughput.

Scenario 4: Terminologically inconsistent output. MT engines often translate the same term six different ways in one document. Correcting terminology inconsistency across 30,000 words isn’t faster with MT - it’s actually harder, because you’re auditing a draft rather than building consistency from the start.

None of these scenarios mean MTPE is a bad workflow. They mean the productivity gain is conditional - and the condition is MT quality, which varies by engine, language pair, and content domain.

Measuring your own throughput: what to track¶

The standard benchmark figures are useful for initial estimates. They’re not useful for tracking what your specific editors, engines, and content types actually produce.

The four metrics worth tracking on every MTPE project:

Words per hour by editor and project type. Not a single average - a breakdown. An editor who does 1,200 words/hour on EN→FR tech docs and 400 words/hour on EN→PL legal docs needs different rate structures for each.

Edit distance (Translation Edit Rate, TER). How much of the MT output does the editor actually change? TER gives you a direct measure of MT quality independent of editor speed. If TER is consistently high (lots of changes), the “post-editing” workflow is actually closer to retranslation.

Effective hourly rate at the project rate. If you’re paying an editor $0.04/word for LPE and they’re doing 400 words/hour, their effective hourly earnings are $16/hour - below professional market rates. That’s not sustainable. Track it and either adjust the rate or the workflow.

First-month productivity vs. steady-state. Separate the ramp-up period from sustainable output. Presenting first-month numbers as benchmarks to management or clients will consistently disappoint.

ISO 18587:2017 - the international standard for MTPE - doesn’t mandate specific throughput figures, but it does require that the post-editing level and quality expectations be agreed in writing before work starts. Having those documented makes it easier to have honest conversations when productivity doesn’t match projected benchmarks.

What 66% faster actually means for an agency¶

The “66% faster” average figure that appears in multiple industry analyses translates to this at scale:

A human translator doing 2,000 words/day costs you $X to produce that volume. A post-editor doing 3,300 words/day (66% more) costs you a meaningful fraction of $X per word (because you pay per word, not per day). At equivalent rates, MTPE expands capacity - you can take on more work without proportional cost increase.

The volume math only works when: 1. MT quality is consistent enough that the throughput is actually 3,300 words/day and not 2,100 2. Post-editor pay scales appropriately so you can retain skilled editors 3. The content type and language pair actually deliver the projected gain

Weglot’s analysis puts the productivity multiplier at 2.5x for structured content in favorable language pairs - matching the upper end of the benchmark table above. That 2.5x is real when conditions are right. It’s not a guaranteed outcome of using any MT engine on any content.

FAQ¶

How many words per hour can a post-editor process?¶

For light post-editing of good MT output: ~1,000 words/hour on average. Full post-editing runs 600-800 words/hour. Human translation from scratch is 250-400 words/hour. These are averages - actual throughput depends heavily on MT quality, language pair, content type, and editor experience with the specific engine.

Is post-editing always faster than translating from scratch?¶

No. SwissGlobal’s 2026 analysis of 90 million words found EN→SV MTPE was 7% slower than human translation. EN→PL showed only an 18% speed advantage. A 2025 Finnish study found individual variation from -2% to +102% across post-editors working on the same content. Poor MT quality, difficult language pairs, and unfamiliar content domains can eliminate the productivity gain entirely.

What content types give the best MTPE productivity?¶

Software UI strings, e-commerce product descriptions, and news articles on predictable topics consistently produce the highest throughput - 1,000-1,500 words/hour for light PE. Legal documents, medical texts, and literary content produce the lowest, often 250-500 words/hour. The difference comes from MT accuracy in specialized domains: MT models trained on general text fail more often on specialized terminology, and every terminology error requires verification time.

How does language pair affect post-editing speed?¶

Dramatically. SwissGlobal’s data shows EN→FR at +130% speed over human translation, EN→PL at +18%, and EN→SV at -7%. High-resource European language pairs with strong MT training data produce the best results. Low-resource pairs, morphologically complex targets, and language pairs with significant structural differences from English show smaller - or negative - productivity gains.

How long does it take a post-editor to reach full productivity?¶

Approximately 3 months of regular work with the same MT engine. Post-editors build familiarity with the engine’s specific error patterns, which allows them to anticipate mistakes and correct them faster rather than re-evaluating each segment fresh. General translation experience transfers partially, but engine-specific familiarity is its own skill.

What is a realistic words-per-day target for MTPE projects?¶

For project planning: light PE at 5,000 words/day (favorable conditions), full PE at 2,000-3,000 words/day. The EU Translation Centre uses 15 pages/day for full MTPE and 20 pages/day for light MTPE as institutional standards. Use these as planning baselines, then run a pilot on your specific content/engine/pair to calibrate for reality.

Sources¶

• Slator - How Fast Can You Post-Edit Machine Translation
• SwissGlobal - Translation productivity 2026: human translation vs MTPE (90M words, 879 linguists)
• Crowdin - Machine Translation Post-Editing Complete Guide
• GTS Translation - The State of MTPE in 2025: What Translators Think (n=212)
• Weglot - MTPE rates and productivity breakdown
• ISO 18587:2017 - Post-editing of machine translation output: requirements