How PDFs, Images, and Files Count Against Claude Usage

Why Files Cost More Than They Look

A file feels like a reference you hand over once. To Claude, it is part of the conversation. Each new reply is produced with the conversation so far as input, which includes the extracted content of your uploads. A 50-page PDF uploaded at message 3 is still in context at message 30.

Anthropic's pricing page puts it directly: how fast you reach a limit depends on the model, how long your chats are, and how large your files are. File size is one of three named factors, which tells you how much it matters.

What the Documentation Says About File Costs

Anthropic does not publish per-file usage costs for claude.ai, but its API documentation describes how the underlying models process files, which is the best available guide to relative cost.

Two takeaways. Visual PDF analysis is the expensive path, because each page carries both text and image cost. And a text-only DOCX of the same content is cheaper than a scanned or image-rich PDF, at the price of losing charts and layout.

Upload Limits Worth Knowing

Per the Claude help center as of November 2026:

• Chat uploads: up to 500MB per file and up to 20 files per chat.
• Images: up to 8000x8000 pixels.
• PDFs: up to 1,000 pages. Larger files are rejected with an "Uploaded file is too large" error.
• Project files: up to 30MB per file, with no cap on file count, but total content must fit within the context window unless RAG applies.
• Supported documents include PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, and XLSX (XLSX needs code execution and file creation enabled).

Images have their own wrinkle. Claude's vision docs explain that images larger than a model's native resolution are downscaled before processing. That caps the cost of a single huge image, but it can also make small text unreadable. A full 4K screenshot of a dashboard is often both more expensive and less legible than a cropped capture of the one chart you care about. If you need Claude to read fine print in an image, crop to it rather than relying on a larger upload.

Upload limits are about what Claude accepts, not what is cheap. A 900-page PDF is accepted and will be read as text, but asking several questions about it in one thread is one of the fastest ways to empty a session.

Projects and RAG: When Files Stop Loading in Full

Project knowledge behaves differently from chat uploads once it gets big. Anthropic's help center says RAG automatically activates when a project approaches or exceeds the context window limit, expanding capacity by up to 10x. Claude then uses a project knowledge search tool to retrieve relevant parts instead of loading everything. RAG for projects is available on paid plans.

This is good for capacity but changes the tradeoffs. Below the threshold, all project knowledge sits in context for every chat. Above it, Claude retrieves pieces, which is efficient but can miss context that a full read would have caught. Either way, irrelevant files make things worse: they add cost below the threshold and add retrieval noise above it.

Anthropic's usage tips include removing project files you no longer use. For structuring a project so it stays lean, see setting up a Claude project knowledge base.

Practical Trimming: A Checklist

1. Extract the relevant pages. If your question is about sections 4 and 7 of a 120-page report, upload those pages. Most PDF tools can split a file in seconds.
2. Prefer text when visuals do not matter. Paste the text of a contract clause rather than uploading the scanned contract.
3. Crop and resize screenshots. Crop to the error message or the relevant UI. Claude's vision docs recommend downsampling when you do not need full fidelity.
4. Pre-filter spreadsheets. Export only the columns and rows you are asking about. See analyzing CSV and spreadsheet data with Claude.
5. One file, one job. Once you have extracted what you need from a document, carry the findings into a fresh chat without the document.
6. Prune project knowledge monthly. Remove superseded drafts, duplicates, and files no chat has used recently.

When not to trim: if you do not yet know where the answer lives, or the task is "find inconsistencies across this whole document," the full file is the point. Pay for it once, ask your best questions early, then move on.

Worked Example: A 120-Page Report

You need to summarize the risk factors and the revenue table from a 120-page annual report.

Full upload: at 120 pages, the file exceeds the 100-page visual analysis threshold, so claude.ai reads text only. At the API docs' 1,500-3,000 tokens per page, that is roughly 180,000-360,000 tokens of text, and you lose the revenue table's layout anyway. Every follow-up carries that weight.

Trimmed upload: extract the 12 risk-factor pages and the 2 pages with the revenue table into a 14-page PDF. It is under the visual threshold, so the table is analyzed visually, and the text load is roughly a tenth of the full file. Then prompt precisely:

Attached: risk factors (pp. 1-12) and the revenue table (pp. 13-14)
from a company's annual report.

1. List the five risk factors most likely to affect revenue in the
   next 12 months, one line each, with the page reference.
2. Reproduce the revenue table as a Markdown table.
3. Flag any risk factor that ties directly to a line in the table.

Do not search the web.

Better answer, smaller context, and the table actually survives. The cost is two minutes in a PDF tool.

Common Mistakes With Files

• Re-uploading the same file in every chat. If you reference a document repeatedly over weeks, a project with that file is cleaner than ten chats with ten copies. If you only need a few facts from it, save those facts as text instead.
• Uploading "just in case." Attaching five background documents to a question about one of them adds load to every reply and gives Claude more material to confuse.
• Keeping a file-heavy thread for unrelated follow-ups. Once the analysis is done, the file is dead weight. New topic, new chat.
• Assuming a DOCX includes its images. The help center notes that for non-PDF documents Claude extracts text only and cannot read embedded images. If the diagram matters, export to PDF or upload the image separately.
• Treating a 900-page PDF as a database. Long PDFs above 100 pages are read as text only. For structured lookups across large documents, a project with RAG or a pre-filtered extract usually works better.

None of these is catastrophic on its own. Together they explain a lot of sessions that end early with only a handful of messages sent.

Where ClaudeKit Fits

ClaudeKit's token counter estimates tokens for what you type and for Claude's responses (roughly within 10%), which helps when you paste extracted text instead of uploading a file: you see the size before you send. The usage circle in the bottom-right of claude.ai shows how much of the session and week a file-heavy thread has taken, on hover, when Claude returns usage data.

Once a document has been analyzed, pin the key findings to the Workspace panel or create a Handoff Brief, then continue in a fresh chat without the file. Export the original thread as Markdown if you want a record of the full analysis.