Cut your AI token bill by up to 70% by turning your prompts into pictures
A proxy renders bulky prompt text as images to dodge token costs, cutting AI bills by up to 70 percent. Here is how the trick works, and where it quietly breaks.
By The Daily Query · · 3 min read
query://opinionHere is an idea that sounds like a joke until you sit with the math for a minute. A developer publishing under the handle teamchong built a tool, pxpipe, that shrinks your AI bill by taking the boring bulk of a prompt and rendering it as a picture. It does not summarize the text or throw any of it away. It screenshots it and sends the image instead of the words, and the model reads the image back.
The reason it works is a pricing asymmetry most people never think about. Text costs tokens roughly in proportion to how many characters you send. An image, though, costs a flat amount based on its pixel dimensions, and it does not care whether those pixels show a blank wall or a wall of dense code. So if you pack text tightly enough into a fixed-size image, the cost per character falls off a cliff. pxpipe's own numbers put it plainly: a 1928 by 1928 pixel image runs about 4,761 vision tokens and holds roughly 92,000 characters, which works out to around 3.1 characters per token against about 1 character per token for plain text.
pxpipe is a local proxy that sits between Claude Code and the API. It watches your requests, spots the parts that are safe to compress, the older conversation history, big tool outputs, static system prompts, renders them to images, and forwards everything on. You start it with one command and point the tool at it. The author reports a 59 to 70 percent cut in end-to-end cost on Fable 5, measured against uncompressed runs rather than a friendly guess.
The catch is spelling
There is a real cost, and the README does not hide it. This kind of compression is lossy for exact strings. When a model reads an image, it does not see letters, it sees patches that it decodes using its sense of what text usually looks like. That is fine for prose, where context fills the gaps. It is a problem for a 12-character hex string, where every character is load-bearing and there is no context to lean on. The tool's own benchmark tells the story: those exact hex strings scored 13 out of 15 on Fable 5 and a brutal 0 out of 15 on Opus. So anything byte-critical, an ID, a hash, an API key, has to stay as text, and pxpipe keeps recent turns uncompressed for the same reason.
Different models, different eyesight
The other wrinkle is that this trick depends entirely on how well a given model reads pictures of text. Fable 5 handles imaged content well enough that pxpipe turns it on by default. Opus 4.8 and GPT 5.5 misread it badly enough that they are opt-in only. I find that detail more interesting than the headline savings. It means the vision half of these models has a quiet, uneven skill at reading rendered text, and a tool like this only pays off where that skill happens to be sharp.
Why I keep thinking about it
Strip away the cleverness and pxpipe is arbitrage on a seam in the pricing. Nobody set vision-token rates with "what if I cram 92,000 characters into one square image" in mind, and someone noticed the gap and drove a truck through it. That is the part I enjoy. It is the same instinct that finds the cheap flight with two layovers, applied to an AI invoice.
It is also a little cursed, and I mean that fondly. You are feeding a language model a photograph of the thing you could have just told it, trusting its eyes over its ears to save money, and accepting that it will occasionally hallucinate a character because a 5 and an S looked alike at that resolution. The honesty is what sells it. The project documents its own failures, publishes the benchmarks that make it look bad, and tells you exactly which data to keep out of the compressor.
Read the fine print, literally
Whether the specific technique lasts is anyone's guess. Providers can reprice vision tokens, and models can get better or worse at reading images from one release to the next, which would move the whole calculation. The durable lesson is smaller and more useful: token pricing has seams, the people paying real money to run AI coding assistants all day will keep finding them, and the smartest tools in this space are the ones honest enough to tell you where they break. A 60 percent discount is worth a bit of weird, as long as you know which characters not to trust to a picture.
enjoyed this one?_
Get the next one in your inbox.
One email every morning. The best AI tools, reviews, and reads, decoded in five minutes.
up_next → Opinion
Gemini 3.5 Pro Is Three Weeks Late, and 'Announced' Has Stopped Meaning Anything
Google announced its flagship model at I/O on May 19. As of mid-June it's still in limited preview, and prediction markets are handicapping the ship date. The AI industry has a vaporware habit.
