Claude can now see the pictures in your PDFs

If you have ever pasted a 40-page financial report into an AI and watched it completely ignore the charts on page 12, Anthropic has some good news.

Anthropic rolled out native PDF support for Claude 3.5 Sonnet in public beta, and it does more than just scrape text. The model can now actually look at the images, charts, and diagrams embedded right inside the file.

Reading the text, seeing the page

Most LLMs read PDFs by running an underlying text extraction script. That works fine until you hit a two-column layout, an annotated diagram, or a complex financial table.

Claude handles this with a hybrid approach. First, it extracts the raw text from the document. Then, it converts each page into an image for visual analysis, combining text and visuals so it can understand both the words and how they are laid out.

This makes it far more capable of translating complex documents, auditing legal forms, or digesting equity research reports. You can also pair this with tool use, meaning Claude can spot a metric inside a chart and immediately pass it into another software workflow.

Specs, pricing, and guardrails

The feature is currently available in Claude Chat feature preview and via the API using the beta header anthropic-beta: pdfs-2024-09-25. Anthropic plans to bring it to Amazon Bedrock and Google Vertex AI later.

There are clear boundaries. Files must stay under 32 MB and 100 pages, and encrypted or password-protected files are not supported. If you have larger files, Anthropic recommends splitting them up.

Cost-wise, Anthropic is not adding an extra fee on top, but it does consume standard tokens. A single page typically eats up between 1,500 and 3,000 tokens depending on density. That means a 50-page document will run through tokens fast, so Anthropic strongly advises using prompt caching if you plan to query the same file repeatedly.

Why it matters

We think this is one of the most practical upgrades Claude has received in months. Real-world business documents rarely look like clean Markdown; they are messy PDFs full of graphs, screenshots, and visual tables.

For analysts, legal teams, and researchers, this removes the need to build clunky custom OCR pipelines just to extract a single chart. It brings Claude closer to Google Gemini 1.5 Pro in native multi-modal document processing, which has been Google's biggest advantage.

That said, the token footprint is heavy. At up to 3,000 tokens per page, feeding an entire 100-page prospectus could cost you roughly 300,000 tokens on the first call. It is a fantastic tool when visuals are crucial, but for text-only manuals, plain text extraction remains the smarter and cheaper bet.