Image Input with Claude

Vision input is useful for explaining a diagram, extracting visible fields or reviewing a screenshot. It can miss small text, misread a chart or infer details that the image does not support.

#Send a local image

The following Python example requires the anthropic package, an authorized API account, ANTHROPIC_MODEL and a local diagram.png. Running it sends image data to the provider and may incur charges.

pythonCopy
import base64
import os
from pathlib import Path
from anthropic import Anthropic

encoded = base64.b64encode(Path("diagram.png").read_bytes()).decode("ascii")
response = Anthropic().messages.create(
    model=os.environ["ANTHROPIC_MODEL"],
    max_tokens=512,
    messages=[{"role": "user", "content": [
        {"type": "image", "source": {
            "type": "base64", "media_type": "image/png", "data": encoded
        }},
        {"type": "text", "text": (
            "Describe the components and arrows in this diagram. "
            "Mark unreadable labels as unreadable; do not guess."
        )},
    ]}],
)
for block in response.content:
    if block.type == "text":
        print(block.text)

The MIME type must match the actual file. Check current supported formats, request size and image-count limits before building upload validation. More pixels do not always improve results after provider-side resizing.

#Make the task verifiable

For a chart, ask for the axis labels and visible values before asking for a trend. For a document, request a source page or region with each field and keep unknown values null. Crop an authorized region if text is unreadable; do not substitute a confident guess.

Validate critical numbers against the source. Remove secrets and unnecessary personal data before upload. Image interpretation is not proof of identity, exact counting or a guarantee of OCR accuracy.

Vision documentation↗