Per-Word Pricing: Examples & Companies

What is it

Per-word pricing is a billing unit common in translation and localization platforms where the metered object is the word count of content processed.

The unit predates the current AI era by decades: human translators and language-services vendors have priced per word for generations because a document’s word count is the most direct proxy for the effort to translate it. Cloud and then AI-assisted platforms inherited it because buyers in localization already tracked words and procurement already had per-word benchmarks — matching that vocabulary removed friction from the sales conversation.

Per-word billing spans two product contexts that share the unit but not the economics. Translation and localization — Crowdin, Lokalise, Unbabel — meter words (under various labels) because localization managers and language-ops leads expect words on an invoice. AI content and voice — Copy.ai for writing, PlayHT for text-to-speech — metered words or their sub-word equivalent, characters, because creators paste scripts of known length. Across both, the word is the atom that makes cost legible to the buyer.

AI now pressures the rate from both sides: as machine-translation quality rises, the marginal cost of a word collapses, forcing platforms to cut rates, shift the meter to a higher-value dimension, or drop per-word billing entirely.

How it works

Per-word pricing has four structural variants in the corpus. The key distinction inside localization is what the word count measures: standing library (hosted) or translation activity (processed).

The hosted/processed split changes the bill for identical content. Hosted words count what you keep: adding a target language multiplies the meter even in a month with zero translation, so the model rewards stable, slow-moving libraries. Processed words count what you move: a large corpus translated once a year stays cheap, but frequent small updates burn the annual pool fast.

Unit math: The same 10,000-word product translated once into 6 languages consumes 60,000 units under either meter — but Crowdin charges those 60,000 as standing hosted words the month you add the languages, while Lokalise draws them from the annual processed-word pool only when the translation runs. Same number, opposite trigger.

At the pure per-unit end, rates track service quality — human-verified translation with a quality SLA sits an order of magnitude above raw machine output (see the FAQ for the bands). That collapse toward fractions of a cent is why several vendors here stopped exposing a per-word rate at all and bundle the economics into subscription tiers instead. PlayHT’s consumer plans charged $4 per additional 10,000 characters once the annual allotment ran out.

Companies using this

Five companies in the corpus meter words as a billing unit, spanning translation management (Crowdin, Lokalise, Unbabel), AI content writing (Copy.ai), and text-to-speech (PlayHT). The table spans the full spectrum from fully transparent self-serve rate cards to sales-only enterprise quoting.

Patterns observed

Vendors keep the word but change what it means. The word is so entrenched that platforms retain it even after their cost structure stops mapping to it — they re-base it rather than drop it. Crowdin moved from per-string billing (used from its 2009 launch through at least 2019) to a hosted-words capacity meter, re-pointing the word at library size. Lokalise, in its 2025 repackaging, moved from hosted keys and seat limits to processed words, re-pointing it at annual throughput. Same familiar unit on the invoice; a different thing being counted underneath.

The word is a good meter only when input length predicts cost. That is why PlayHT metered the text going in — creators paste scripts of known length — rather than the audio coming out. The choice held for consumer content but became a liability as the product moved toward voice agents, where identical character counts produce widely different speech lengths. The meter that matches the buyer’s mental model can still be the wrong meter for the cost being incurred.

Counterexamples & variants

Unbabel drops the per-word rate at the enterprise level. Its core LangOps platform publishes no per-word figure — the /pricing URL 404s and the only buying surface is a sales-contact form — because blended AI-plus-human translation makes cost per word swing with the quality SLA, language pair, and content type, so a single headline rate would misrepresent it. Its self-serve sibling Widn.ai instead publishes flat monthly tiers (Free, $19/mo, $90/mo, Custom): even a vendor whose obvious value metric is the word reaches for a subscription the moment it sells self-serve.

Lokalise’s processed-words meter penalizes velocity. It rewards large-but-infrequent corpora — a 5M-word product translated once a year fits an annual allowance — but punishes agile teams shipping small frequent updates across many languages. The Growth plan ($375/mo) includes from 225K processed words/year; a 12-language, 3,000-word update run (36,000 processed words each) drains that in roughly six cycles, pushing high-velocity teams onto the Advanced tier ($999/mo) for capacity, not features. An allowance sized for project-based translation is a structural mismatch for continuous, small-batch localization.

What this means for buyers vs vendors

For buyers

Pin down which word the vendor counts before you compare rates — source, hosted, or processed imply different bills for identical content. Then probe whether the count includes whitespace, markup, and regenerations, all of which inflate the meter past the visible document length. If the vendor runs both AI and human-review tiers, benchmark them separately: the word count is the same, but a human-reviewed word can cost an order of magnitude more than a machine-translated one. For a framework on evaluating usage meters, see choosing the right usage metric.

For vendors

Per-word pricing is a credibility signal in localization — buyers expect it and it shortens procurement — but as marginal word-processing cost approaches zero, a per-word rate either becomes too low to capture value or stays anchored to human-equivalent rates your cost structure no longer justifies. The strongest evolution in the corpus carves a separate premium Pro AI word pool alongside the base meter, keeping the vocabulary buyers understand while introducing a value-aligned dimension. The sharper lesson: if the word is your cost driver but workflow automation is your value driver, price the workflow, not the word. See introduction to usage-based pricing for value-metric selection from first principles, and model metering approaches with the usage-based pricing calculator.

Explore this theme in the knowledge graph →