pxpipe cuts token cost by rendering code to images for model OCR

Category: Agentic coding
Status: discussion
Sources: pxpipe repository, HN discussion
Summary: pxpipe is a single-author project that renders source code to images and has the model read it back by OCR, with the author reporting about a 60 percent reduction in Fable token cost. The technique relies on image tokens encoding more characters than text tokens for the same content.
Comments: HN commenters note the approach mirrors a DeepSeek write-up on text-as-image tokens and earlier experiments that saved prompt tokens but needed more completion tokens and ran slower and more expensive overall. Several argued it exploits a provider billing quirk (some backends OCR PDFs internally without charging for the recovered text) that would disappear if pricing were adjusted, rather than a durable efficiency gain.
Why it matters: It highlights how current model pricing across input modalities can be arbitraged, and how fragile such savings are to a provider changing how it meters image versus text tokens.