Can Claude Translate Documents and PDFs? Capabilities in 2026

Your client sent a 47-page contract as a PDF and needs a translation by tomorrow morning. You open Claude, drop the file into the chat, and type “translate this to English.” Will it work? Let’s break down what Claude actually does with documents, where the real limits are, and when you’d be better off reaching for a different tool.

What Claude actually does with documents¶

Claude isn’t just a chatbot that reads and responds to text - it genuinely processes uploaded documents. When you upload a PDF or DOCX, Claude handles it two ways simultaneously: it extracts the text layer and analyzes each page as an image. This means it works with not just the body text but also tables, diagrams, and charts embedded in the file.

For translation, that means concrete things:

• Upload a 30-page contract PDF → ask Claude to translate from English to German → get the translation right in the chat
• Upload a DOCX with a complex table → Claude translates both the column headers and the cell contents
• Upload a technical manual → Claude keeps terminology consistent throughout the document, because it sees the whole thing at once

That last point is a real advantage over older approaches. Thanks to a large context window (200,000 tokens for Sonnet, up to 1 million tokens in the Opus 4.6 beta), Claude holds the entire document in context and doesn’t “forget” how it translated a term on page 5 when it gets to page 40.

What file formats Claude supports¶

Through claude.ai, you can upload files in these formats:

Upload limits through claude.ai: - Maximum file size per file: 30 MB - Maximum files per conversation: 20 files - For PDFs: up to 100 pages (on standard Sonnet models with 200K context)

Important: if a PDF is password-protected or encrypted, Claude can’t read it at all. You’ll need to remove the protection first.

There’s also a less obvious technical constraint: a large PDF with dense content (tiny font, complex tables, heavy graphics) can fill the context window before Claude finishes the document - even if it’s technically within the page limit. For documents over 50 pages, splitting by section gives more reliable results.

Translation quality: what the benchmarks show¶

There are real numbers here, not marketing claims.

At WMT24 - the industry’s primary annual machine translation benchmark competition - Claude 3.5 ranked first in 9 out of 11 language pairs. For German, Japanese, Korean, Dutch, and Italian, Claude was at the top.

In 2025, Lokalise ran a blind study where professional translators evaluated translation quality without knowing which AI generated it. Result: 78% of Claude’s translations rated “good” - the highest score among all tested AI models, ahead of GPT-4, DeepL, and Google Translate.

On MachineTranslation.com’s internal benchmarks, Claude scores 93.8 out of 100 on translation quality.

That said, there’s a meaningful gap by language pair. As MachineTranslation.com notes:

Claude performs strongest with high-resource languages including English, Spanish, French, German, Portuguese, Japanese, Chinese, and Korean - achieving 88-92% accuracy on these pairs. Lower-resource languages show meaningful performance drops, with Southeast Asian and African languages reaching only 72-78% accuracy.

For English-German, English-Spanish, English-French translations - Claude is very strong. For less common language pairs, the quality is lower and you should review the output.

One more thing that doesn’t get talked about enough: Claude can translate confidently and still be wrong. This is most common with specialized terminology - legal, medical, technical. The translation will look natural and sound plausible, but the specific term might be off. For documents that matter, always have someone with subject-matter expertise review the output.

Where Claude falls short¶

There are specific cases where Claude is simply the wrong tool:

Low-quality scanned documents. Claude processes scanned PDFs through its vision module (it looks at the document as an image). If the scan is clear and straight - no problem. But if the document was photographed at an angle, is blurry, has poor contrast, or has handwritten text on top - translation quality drops sharply. Claude can misread characters and won’t always realize it happened.

Encrypted or password-protected PDFs. If the file has a password, Claude simply can’t open it.

Documents with complex layouts. If a document has an unusual design - multiple columns, text layered over images, non-linear text blocks - Claude can mix up the text order or miss sections.

Very old or poorly scanned documents. Soviet-era archival documents, old typewritten pages, historical manuscripts - where even a human needs to work hard to read the text, Claude will make more errors.

Legal translation that needs official validity. This is the most important limitation. As legal experts point out directly:

An AI engine cannot be held legally accountable, nor can it provide the certification required by state authorities. Only sworn translators registered with the Boards of Trade can perform this work.

For embassies, courts, immigration offices, universities, notaries - anywhere that requires a “certified translation” or “sworn translation” - a Claude translation has no legal standing. You need a human translator with the right credentials and an official stamp. Claude can produce a solid draft for a certified translator to review, but it can’t replace the translator.

For more on why official institutions reject AI translations, see Why USCIS, IRCC, and UKVI Reject AI Translations.

Scanned PDFs: a separate conversation¶

This case comes up a lot because there’s genuine confusion about it. Claude really can read scanned documents - it sees the PDF as an image and attempts to extract the text. But there’s a meaningful difference between:

Good scan - scanned on an office scanner at 300+ dpi, black text on white background, pages straight. Claude will handle this well.

Poor scan - photographed on a smartphone, tilted, with shadows, uneven lighting, handwritten corrections over the printed text. Problems start here.

Practical rule of thumb: if you can easily read all the text in the scan yourself, Claude will manage fine. If there are parts where you’re guessing what it says, Claude will make even more mistakes.

Two practical approaches for better results with scanned files: 1. Pre-process the scan through an OCR service (Adobe Acrobat, ABBYY, or even Google Drive) - it converts the image to a text-layer PDF 2. Copy the recognized text manually and give Claude that text to translate

The second option gives cleaner results when the scan is complicated.

Practical tips for better translation output¶

If you’re already using Claude for documents, a few techniques make a noticeable difference.

Give context and register upfront. Instead of “translate this,” try: “Translate this legal contract from English to French. Keep the formal legal register and preserve legal terminology. If a term has no direct equivalent, leave the original in brackets.” The details in the prompt matter.

Set terminology in advance. If you know how specific terms should be translated, say so at the start: “In this document, ‘beneficial owner’ should always be translated as ‘bénéficiaire effectif’.” Claude will maintain this throughout the conversation.

Translate in sections for long documents. For documents over 50 pages, upload and translate by chapter or section. It keeps quality consistent and lets you catch problems before you’re 45 pages in.

Be specific about formatting. If you only need the text without structure, say “translate the text only, no formatting.” If you need structure preserved, say “keep all headings and numbering.”

Spot-check key terms. After translation, ask Claude: “Check whether the translation consistently uses the same term for X throughout.” This catches inconsistencies, especially in long documents.

For more on prompting strategies for translation - see Prompt Engineering for Translation: Getting Better Results from ChatGPT and Claude.

Claude vs other AI tools for document translation¶

A quick comparison of the main options:

For a detailed quality comparison with benchmarks - see Claude vs ChatGPT for Translation: What the Benchmarks Show in 2026.

If you’re concerned about confidentiality when uploading documents to AI services - that’s an important separate topic. Free-tier Claude plans store conversations and may use them for model training. For details, see AI Translation and Privacy: What Happens to Your Data.

FAQ¶

Can Claude translate PDF files?¶

Yes. Upload a PDF up to 30 MB (or up to 100 pages) directly to claude.ai, ask for a translation, and get the result in chat. Claude reads text, tables, and diagram captions. Encrypted PDFs aren’t supported.

How does Claude’s translation compare to DeepL?¶

In the Lokalise 2025 blind study, Claude received the highest ratings from professional translators: 78% of translations rated “good.” DeepL is traditionally strong for formal business text, but Claude has the edge in flexibility and handling long documents where terminology consistency matters. For most tasks they’re comparable in quality, but Claude handles multi-section documents more coherently.

How many pages of PDF can Claude translate at once?¶

Up to 100 pages through claude.ai on standard models. Dense text (small font, complex tables, lots of graphics) can fill the context window before that limit. For documents over 50 pages, translating by section is more reliable.

Can Claude translate Word documents (DOCX)?¶

Yes, DOCX is natively supported. Claude translates both the body text and tables inside the document. It delivers the result as text in the chat - if you need a DOCX output, you’d need to copy it back into a Word file or ask Claude to format the output as markdown for easier conversion.

Does Claude store my uploaded documents?¶

Depends on the plan. Free and Pro plans: conversations are stored and may be used for model training (you can disable this in privacy settings). Claude API: data is never used for training, logs are deleted after 7 days. For client confidential documents, either use the API version or turn off training in your account settings.

Can Claude translate handwritten documents?¶

Partially. If the handwriting is clear and well-scanned, Claude does reasonably well. Unclear handwriting, old-style cursive, historical scripts - errors start creeping in, sometimes significantly. For important handwritten documents, a specialized OCR solution will do better.

Is a Claude translation accepted by embassies or government agencies?¶

No. Official submissions require certified translation from a sworn translator or a notarized translation. A translation from any AI - Claude, ChatGPT, DeepL - has no legal validity for government institutions. Claude can produce a solid draft, but the final version needs to be reviewed and signed by someone with the appropriate legal credentials.

Sources¶

1. WMT24 Machine Translation Benchmark - official benchmark competition results
2. MachineTranslation.com - How good is Claude for translation - quality analysis and benchmarks
3. Anthropic PDF Support Documentation - official documentation on PDF handling
4. Leaders League - AI risks for legal translation - AI risks in legal translation
5. Claude Pricing 2025 - DataStudios - pricing and usage limits