Vision via API
Vision via API is Claude's capability to accept and reason about image inputs alongside text within the standard Messages API. Developers send images as base64-encoded data, publicly accessible URLs, or file references (via the Files API), and Claude returns text-based analysis, descriptions, or structured data about what it sees. No separate endpoint or special configuration is required — images are simply inserted as content blocks in the same request structure used for ordinary text conversations.
Under the hood, Claude processes images through the same reasoning architecture it uses for language, which means it can interpret charts, extract text through OCR, compare visual states across multiple images, and reason about semantic relationships in visual data — not just classify objects. This distinguishes it from narrow computer vision pipelines optimized for a single task.
All current Claude models (Haiku, Sonnet, and Opus families) support vision. Images are counted as tokens, so visual input affects cost and context window usage. The API imposes a 32 MB per-request payload limit and supports JPEG, PNG, GIF (first frame only), and WebP formats. Up to 600 images can be included in a single request when using the full 1M-token context window.
When you’d use it
- ◆Document and Receipt OCR — Extract structured data from invoices, receipts, contracts, or handwritten forms without building a custom OCR pipeline. Claude reads the text, understands context, and returns formatted output ready for database insertion or expense reporting.
- ◆Chart and Graph Analysis — Upload a screenshot of a sales dashboard, financial chart, or analytics report. Claude identifies axes, trends, anomalies, and key data points, then summarizes findings or returns structured JSON for downstream processing.
- ◆Quality Assurance and Defect Detection — Submit a reference image of an acceptable product alongside a real-time photo from an assembly line. Claude compares the two, identifies missing components or visual discrepancies, and flags items for human review — without requiring specialized model training.
- ◆Agentic Screen Navigation and GUI Automation — Feed sequential screenshots of a desktop or web browser to an autonomous agent. Claude interprets the visual layout, identifies buttons and fields, and generates navigation steps or coordinate-based actions to fulfill a natural language command — enabling robotic process automation.
- ◆Multi-Image Compliance Auditing — Submit multiple inspection or architectural photos alongside reference criteria. Claude reviews each image, checks for compliance issues, cites relevant codes or standards, and produces a prioritized remediation report with estimated costs.
What changed recently
- ◆2026-05 — Claude Opus 4.8 released with improvements to agentic vision and computer use benchmarks (agentic computer use score reached 83.4%), making it better at navigating GUIs and interpreting sequential screenshots in autonomous workflows.
- ◆2026-04 — Claude Opus 4.7 released with high-resolution image support: maximum image resolution increased to 2576px on the long edge (up from 1568px on prior models). High-resolution images can use approximately 3x more tokens (up to ~4784 tokens per image vs ~1568 previously). Up to 600 images per request supported with the 1M token context window.
- ◆2026-04 — When more than 20 images are submitted in a single API request, the maximum dimensions per image are reduced to 2000x2000px (down from 8000x8000px for requests with 20 or fewer images).
This is the short version
The full chapter has three worked examples, the common pitfalls, and the workflow that makes it pay — plus the other 84 features, kept current.
Get Claude Master — $97 →