Vision (Image Input)

Claude can read images directly. Use for OCR, chart interpretation, screenshot debugging, document understanding, etc.

Sending an image

Two delivery modes:

Base64:

{
  "type": "image",
  "source": {
    "type": "base64",
    "media_type": "image/png",
    "data": base64_str,
  },
}

URL (where supported by the platform):

{
  "type": "image",
  "source": {"type": "url", "url": "https://..."},
}

Up to ~20 images per request typical. Each image consumes tokens proportional to dimensions (~1.15 tokens per pixel-pair).

Resize before sending. A 1568px-on-longest-edge resolution is usually plenty. Larger costs more tokens with no quality gain.
Mention what to look at. "In the screenshot, identify the error in the second log entry" beats "what's wrong?"
For multiple images, label them. "Image 1 shows X, image 2 shows Y. Compare..."
For OCR-heavy work, use PDF input (document blocks) instead — Claude reads PDFs natively.

{
  "type": "document",
  "source": {"type": "base64", "media_type": "application/pdf", "data": ...},
}

PDFs preserve text + vision — Claude sees both the rendered page and the extracted text.