benchrAll capabilities Documents and files
Read a 600-page PDF including its charts
Not OCR. The model reads the text and looks at each page, so a figure that only exists as a chart is still answerable.
DocsOfficially documentedTestNot tested by benchr
Claude APIclaude.ai and the Claude appsClaude Opus 5Claude Sonnet 5+2
Overview
- Accepts standard PDFs up to 32 MB per request and up to 600 pages, dropping to 100 pages when the model's context window is under 1M tokens.
- Reads tables and charts as images alongside the extracted text, which is what makes financial and legal documents workable.
- Works from a Files API file_id, so the same document can be referenced repeatedly without re-uploading it.
How it works
- 01Upload once through the Files API and reference the file_id, rather than sending base64 on every turn.
- 02Ask page-anchored questions. Requesting the page number with each answer makes verification cheap.
- 03Split anything past the page ceiling into ranges and keep the ranges stable, so answers stay comparable.
- 04Pair it with a JSON schema when the output is going into a database rather than a person's eyes.
Prompt
Read the attached document. For every figure you report, give the page number it came from and say whether it was printed as text or read from a chart or table image. If a figure is not stated in the document, write "not stated" rather than estimating it.The page-and-provenance requirement is what turns a summary into something you can check.
Limits
- Password-protected or encrypted PDFs are out - the requirement is a standard PDF.
- Both the 32 MB and the page ceiling apply to the entire request payload, not just the PDF.
- Above 20 image or document blocks in one request, a stricter per-image dimension limit applies to all of them.
- Scanned pages inherit every vision limitation, including unreliable counting and approximate localisation.
Verification
benchr test runs
benchr has not run this capability. No test has been published.
benchr test runs →
Evidence
Sources
- Anthropic - PDF support: 32 MB, 600 pages, standard-PDF requirement Tier 1 · September 1, 2026
Not stated by the source
- Extraction accuracy on any specific document class
- Whether page count or page complexity drives token cost more
Source-check history
- September 1, 2026DocsOfficially documentedFirst logged. Size, page and format requirements read from the vendor's support page.
Changes
Provider changes affecting this
- August 31, 2026Claude Sonnet 5Pricing changed
- July 3, 2026Claude Sonnet 5Pricing changed
- July 1, 2026Claude Sonnet 5Model released
- June 1, 2026Claude Opus 4.8Model released