July 3, 2026
Screenshot your prompt, save the drama
60% Fable cost cut by converting code to images and having the model OCR it
AI coders are turning giant blocks of text into screenshots to slash bills — and commenters are calling it genius, cursed, and probably a loophole
TLDR: A developer says turning huge AI prompts into images can cut costs by about 60%, but the community is split on whether it’s brilliant optimization or a cheap loophole that will soon be closed. Commenters also warn it may be slower, less reliable, and hilariously painful to look at.
A new coding tool called pxpipe has the internet doing a full double take: instead of sending huge chunks of text to an AI model, it turns that text into images and lets the model read them like screenshots. The pitch is wild but simple — if the AI company charges less for images than for the same words as text, you can cut the bill hard. In the demo, the creator says that can mean roughly 60% cheaper requests, especially when the prompt is packed with code, instructions, and old chat history.
But the comments? Absolute popcorn material. One camp is impressed that this “stupid-smart” trick works at all, with one commenter noting there’s even a DeepSeek whitepaper on similar ideas, which instantly gave the whole thing “forbidden academic hack” energy. The other camp is already sounding the alarm: critics called it a pricing hack that burns extra computer power and predicted the loophole will get shut the second companies notice. Another commenter came in with the classic “been there, tried that” flex, saying they tested a similar idea last year and ended up with slower responses and bigger output costs.
And then there was the pure comedy relief: “Ahhh my eyes the vibe coded readme” became the thread’s unofficial reaction meme, perfectly capturing the chaotic look of giant walls of text mashed into an image. Underneath the jokes is a real warning: this trick can misread exact numbers, IDs, and other tiny details. So yes, commenters think it’s clever — but also possibly cursed, temporary, and one pricing update away from death.
Key Points
- •The article presents pxpipe as a local proxy that converts dense Claude Code request content into images to reduce input token counts.
- •It reports measured efficiency of about 3.1 characters per image token versus about 1 character per text token on real Claude Code traffic.
- •One cited example reduced roughly 25,000 text tokens to about 2,700 image tokens, with estimated end-to-end cost reductions of about 59% to 70% at current Fable prices.
- •pxpipe only compresses requests, not model outputs; recent turns stay as text while system prompts, tool docs, and older bulk history are imaged.
- •The article states the method is lossy and unreliable for exact retrieval, citing needle-in-haystack results of 0/15 on Opus and 13/15 on Fable 5 for exact 12-character hex strings inside imaged content.