Anthropic Updates Claude "Zoom Tool" Cookbook for High-Resolution Image Analysis
Anthropic has updated its "zoom tool" cookbook, enabling Claude to request and analyze high-resolution cropped regions of large images. This prevents the loss of important details that typically occurs when large images are downscaled for model input. Vision LLMs often struggle with fine details in large charts, documents, or diagrams due to automatic downscaling. This tool-use pattern enables more accurate and detailed visual analysis without requiring massive context windows or processing overhead for the entire high-res image. The cookbook demonstrates how to implement a crop tool using Claude's tool-calling capabilities, enabling the model to dynamically request specific coordinates of an image. The system then crops the original high-resolution image at those coordinates and sends the crop back to Claude for detailed inspection.
## BACKGROUND
Multimodal LLMs process images by converting them into tokens, but they often downscale large images to fit within token limits, which can erase small text or fine details. Tool calling, or function calling, is a mechanism that allows LLMs to interact with external code or APIs to perform specific tasks. By combining these concepts, developers can give LLMs the ability to "zoom in" on specific parts of an image programmatically.