Claude Opus 4.7 Vision Evidence OCR Deposition Photos

Claude Opus 4.7 vision accepts images up to 2,576 pixels and 3.75 megapixels — a 3.26x improvement over 4.6's 1.15 MP ceiling. Anthropic shipped this on April 16, 2026 alongside task budgets and the new "xhigh" effort level, per the What's New in Claude Opus 4.7 docs. Most coverage of the vision bump focused on coding workflows and screenshot-to-code use cases. The legal angle the dev press skipped: evidence review, OCR on scanned discovery, and whiteboard capture from depositions all just got materially more reliable. Here's what the resolution improvement actually changes for litigation and how to use it without creating evidentiary problems.

What 3.75 megapixels means for legal documents

A standard 8.5x11 page scanned at 300 DPI is roughly 8.4 megapixels. A page scanned at 200 DPI — common for older firm scans — is about 3.7 megapixels. A page scanned at 150 DPI is roughly 2.1 megapixels.

Claude 4.6's 1.15 MP cap meant most legal scans had to be downsampled before upload. The downsampling lost detail at exactly the wrong places: small footnote text, redaction artifacts, faint signatures, marginalia. Claude could read the body of the document but missed the parts a litigator needed.

4.7's 3.75 MP ceiling clears 200 DPI scans without downsampling. 300 DPI scans still need either tiling or single-page processing, but the loss is minimal at that density. For practical discovery work, 4.7 reads scanned documents at the resolution most firms actually scan at.

The second-order effect: paralegals who previously hand-typed key passages from low-quality scans into Claude can now upload the scan directly. Per-document handling time drops materially. The third-order effect: the gap between "AI can review native digital documents" and "AI can review scanned documents" closes for the first time. Most matter-relevant evidence in 2026 is still in PDF form; this is where the legal-AI productivity gain lives. The Opus 4.7 anchor covers the broader change set.

OCR on scanned discovery: where 4.7 wins and where it doesn't

Claude is not a dedicated OCR engine. It's a multimodal model that happens to read images. For OCR on scanned discovery, that distinction matters:

Where 4.7 wins:

- Mixed-content pages; text plus tables plus signatures plus stamps. Dedicated OCR struggles with mixed content; Claude reads the page as a whole and returns structured output. - Document classification; "is this an invoice, a contract amendment, or a meeting note?" Claude classifies in one pass. - Selective extraction; "pull every dollar amount and the surrounding context" or "find all dates with their associated parties." - Multi-language pages; Claude handles French, Spanish, German, Mandarin, Japanese in a single image without language detection overhead.

Where dedicated OCR still wins:

- Bulk extraction; running 50,000 pages through pure OCR is faster and cheaper than Claude vision per page. - Field-level structured output; invoice processing where you need exact line-item alignment. - Regulatory forms; purpose-built form extraction (1099, W-2, IRS forms) is more reliable through dedicated tooling.

The practical workflow: dedicated OCR for bulk text extraction; Claude vision for the documents that matter, the edge cases, and the structured analysis on top of OCR output. The task budgets discovery spoke covers the cost-control side.

Deposition whiteboard capture and visual exhibits

Modern depositions involve more visual content than the 2010s playbook anticipated. Witnesses sketch on whiteboards. Counsel marks up exhibits with highlighters. Reporters now capture these with phone cameras and reference them in transcripts.

With 4.7's 3.75 MP vision input, a phone camera capture of a whiteboard sketch is processable in full resolution. Claude reads the sketch, transcribes any text, identifies the structural elements, and can answer questions about it later in the matter. For deposition prep; "what did Witness X draw on the whiteboard during the December deposition?"; the model now serves as a searchable index of visual content the transcript misses.

The operational caveat: deposition exhibits are evidence. Claude's analysis of a deposition exhibit doesn't carry evidentiary weight; the exhibit itself does. Use Claude vision for prep, search, and analysis, never as a substitute for the actual exhibit in the trial record. Document any AI-assisted analysis of deposition exhibits in your privilege log if asked.

For multi-day depositions, pair the vision capability with multi-session memory. Claude reads a Day 1 whiteboard, the scratchpad captures the analysis, and Day 5 questioning can reference Day 1 visual evidence without re-uploading.

Evidence handling: what to load and what to keep separate

Three rules that should govern AI vision use on evidence:

1. Don't run privileged client documents through consumer Claude. *United States v. Heppner* (SDNY, Feb 17, 2026) ruled consumer Claude exchanges aren't privileged. Use claude.ai Team/Enterprise, the API, AWS Bedrock, Vertex AI, or Microsoft Foundry for any matter-context vision work. The cybersecurity safeguards privileged context spoke covers the deployment-surface decision.

2. Don't replace the original. Evidence integrity requires the unmodified original document or image. Claude vision analysis is *derivative work product*, not the evidence itself. Maintain the chain of custody for the source image; treat Claude's output as analysis to verify against the original.

3. Document AI-assisted analysis where required. Federal court standing orders increasingly require disclosure of AI use in litigation work. Per the Ropes & Gray AI Court Order Tracker, more than 300 federal judges have AI-related standing orders or local rules as of 2026. Some require tool name; some require which sections of work product were drafted with AI assistance. Check the court's standing orders before any matter-context AI vision use.

For counsel handling potentially AI-generated evidence (deepfakes, synthetic exhibits), Claude vision is also a useful first-pass authentication layer. Anomalies that Claude flags often warrant deeper forensic review, but Claude is not a forensic tool; its output is investigative, not evidentiary.

Photo evidence in personal injury and criminal practice

Two practice areas where the vision improvement compounds:

Personal injury. Accident scene photos, vehicle damage documentation, injury photos at intake and over time, surveillance video frames. With 3.75 MP, a phone camera shot of vehicle damage is processable in full detail. Claude can identify damage patterns, compare pre-incident and post-incident photos for inconsistency, and surface details a paralegal might miss on first review. For volume PI practices, this is meaningful intake-stage triage capability.

Criminal practice. Surveillance video stills, crime scene photos, body-worn camera frames, evidence photos. Defense practice especially benefits from vision-assisted review of disclosed video evidence; the prosecution typically discloses thousands of frames and the defense lacks the bandwidth to review each. Claude can summarize patterns, flag specific frames for human review, and answer questions about visual content.

The constraint in both areas: Claude's output is investigative input, not expert testimony. A PI matter still needs an accident reconstructionist; a criminal matter still needs a forensic video analyst. Claude vision accelerates the work that goes into selecting which exhibits matter; it doesn't replace the experts who testify about them. The creative writing brief drafting spoke covers a parallel writing-side workflow.

Cost and procurement implications of vision-heavy use

Image inputs consume tokens. Per Anthropic's documentation, Opus 4.7 image input pricing follows the same $5 per million input token rate as text, with image content tokenized based on resolution and aspect ratio. A 3.75 MP image at typical aspect ratio runs in the low thousands of tokens.

For a litigation practice running 500-1,000 vision-input requests per month at moderate complexity, the additional input-token spend is in the low hundreds of dollars. That's small relative to the analyst hours saved on document review and exhibit handling.

The procurement implication: vision-heavy practices benefit most from Anthropic Enterprise consumption deals where volume commits unlock better unit economics. For mid-market firms on Claude Team at $25/user/month, the vision capacity is included up to plan limits. For BigLaw running custom enterprise deals, negotiate vision-input volume in the renewal terms.

For the consumption-cost analysis on consumption-based pricing, the tokenizer cost calculator lets you model your specific image-input volume against expected token consumption.

The Bottom Line: The verdict: 4.7's vision bump from 1.15 MP to 3.75 MP is a real productivity unlock for litigation, especially for scanned-discovery review, deposition whiteboard capture, and personal-injury photo triage. It's not a replacement for dedicated OCR on bulk extraction or for forensic experts on evidentiary analysis. Use Claude vision for the documents that matter and the edge cases; keep dedicated tooling for high-volume structured extraction.