60% Fable cost cut by converting code to images and having the model OCR it (github.com)

A new technique called "pxpipe" has been introduced to significantly reduce costs associated with using the Claude Code AI model by transforming large text inputs into images. This innovative approach compresses dense content like system prompts and tool documentation into compact PNGs, resulting in a dramatic reduction of input tokens—from approximately 25,000 text tokens to about 2,700 image tokens—leading to end-to-end savings of 59% to 70% depending on the workload. Notably, images have a fixed token cost based solely on their pixel dimensions rather than text density, making this method especially advantageous for token-dense information such as code and JSON. The implications for the AI/ML community are profound, as pxpipe allows users to handle extensive contextual information more efficiently without compromising the quality of responses. While the images are lossy—meaning they may not perfectly capture exact string values—recent turns remain text-based to preserve accuracy where necessary. As the demand for processing complex queries and large datasets increases, this technique offers a cost-effective way to enhance AI capabilities while maintaining performance, emphasizing the importance of token management in optimizing AI operational expenses.