Translation memory with DeepL, and how the credits work

Memory first, DeepL for the rest

Pre-translating a project in Lexora runs in this order:

1. Exact matches from your memory. Free, always.
2. Fuzzy matches, if you ask for them. Choose the lowest match to insert: 95%, 85%, 75% or 50%. Also free.
3. DeepL or Lexora AI for what is still empty. Tick “AI translation for the rest” and choose the engine. Before anything is sent, the window says how many segments are left and how many credits they will take.

Everything comes in as a draft for you to check. Segments that already have text are never changed, locked segments are left alone, and a draft that would lose a formatting tag is skipped. AI is off until you turn it on for a project, so for a client who doesn’t allow machine translation, nothing leaves your computer. The help article on AI translation shows where each button is.

How the credits are counted

Every plan has one balance of AI credits, used by two engines:

• Lexora AI takes 1 credit for each word of the source, at least 1 a segment.
• DeepL takes 2 credits for each character of the source, spaces and punctuation included. Formatting tags are not counted.

Three sentences, counted by the same code the Lexora server uses:

1. English: “Press the red button to stop the machine.”
2. German: “Drücken Sie die rote Taste, um die Maschine anzuhalten.”
3. English: a 24-word sentence from our translation memory software page, “Before you start, Lexora counts the project’s words by their best memory match, the basis most clients use for quotes, and counts repetitions apart.”

The three sentences have 8, 9 and 24 words, so DeepL took 10 to 12 credits a word, against 1 for Lexora AI. Languages with longer words take more DeepL credits per word.

Why DeepL is counted by the character

DeepL sells its API by the character. On 16 October 2026 its pricing page listed a Growth plan with 12 million characters a year and $27.50 for each extra million characters; the page showed us prices in US dollars (DeepL API pricing). Counting Lexora’s credits the same way means a long sentence costs what it costs us, whatever the language.

One part is free on both sides. When you look up a word or a phrase in the DeepL panel beside a segment, Lexora sends the sentence it comes from as context, and DeepL’s documentation says that “Characters included in the context parameter will not be counted toward billing” (DeepL API, Translate text). You pay credits for the word or phrase, not for the sentence around it.

What a month of credits covers

At the 10 to 12 credits a word of the sentences above, a Pro month spent only on DeepL covers about 32,000 to 39,000 English words of source text. Words your memory fills cost nothing, so a client with a large memory leaves more for the rest. Credits renew every month and don’t carry over (how AI credits work).

What DeepL gets, and what it doesn’t

• It gets the source text of the segments you send, with their tags, and the language pair. The privacy policy lists exactly what goes where.
• It doesn’t get your glossary. DeepL’s API can use glossaries of its own, but Lexora doesn’t send yours to it. Lexora AI does get the matching glossary terms. With either engine, the term check flags a draft that misses a glossary term.
• It doesn’t learn from your corrections. Lexora AI is sent up to eight translations you recently confirmed in the project, with the draft you corrected, so it can follow your choices. DeepL is not.

So we think DeepL is worth its higher credit cost when you want its style for a language pair you know it handles well, and Lexora AI is the better fit for long jobs, for projects with a glossary, and when you want the drafts to follow your earlier corrections. That is our view, and the drafts are yours to check either way.

Read next

• What a translation memory is and how to start using one
• How Lexora’s translation memory software keeps a memory for each client
• Turning AI on for a project, pre-translating and AI review