Fable cuts 60% of costs by converting code to images and using OCR

The open‑source project pxpipe posted on GitHub demonstrates a cost‑cutting pipeline for Fable’s LLM workflow. By rendering source files as PNG images and feeding those images through an OCR engine before invoking the language model, the author measured a 60 % reduction in API spend for a representative codebase [GitHub].

How the pipeline works

pxpipe first converts each source file into a series of raster images, one per line of code. An OCR tool (the repo defaults to Tesseract) extracts the textual content from each image, which is then supplied to the LLM. Because the model receives a single token per image rather than a stream of characters, the token count drops dramatically, and the associated compute bill follows suit. The repository includes a ready‑to‑run script that automates image generation, OCR processing, and LLM invocation, making the approach reproducible for any Python or JavaScript project.

Why it matters

The 60 % cost saving lowers the barrier for smaller teams to experiment with large language models, expanding access beyond well‑funded enterprises. Turning code into images also sidesteps language‑specific tokenization quirks, which could enable more uniform handling of multi‑language codebases. Finally, the project’s permissive MIT license invites community contributions, meaning further refinements—such as faster OCR back‑ends or tighter integration with vision‑augmented models—can be built on top of the existing proof‑of‑concept.

The hack challenges the assumption that text‑only pipelines are the only efficient path for LLM‑driven code analysis, and the early results suggest that visual preprocessing may become a viable cost‑optimization strategy for the broader AI tooling ecosystem.