This open-source tool hides text in images to cut Claude's bill by up to 70%

Here’s a trick that sounds like it shouldn’t work: a new open-source tool called pxpipe takes long text prompts, renders them as PNG images, and sends those to Claude instead — cutting costs by 59 to 70 percent, according to its developer.

The loophole is in how Anthropic prices things. Text costs roughly one token — the small chunks of text AI models bill by — per character. Images cost a fixed number of tokens based on their pixel size, no matter how much text is crammed into them. Render dense content like code or documentation as a tightly packed image, and you fit about 3.1 characters into every image token. In one example, around 48,000 characters of system prompt and tool documentation would cost about 25,000 tokens as text. As a single PNG page: roughly 2,700.

pxpipe, built by developer Steven Chong, runs as a local proxy — a small program that sits between Claude Code and Anthropic’s servers. It intercepts requests and converts the bulky, unchanging parts (system prompts, tool docs, older chat history) into images, while recent messages pass through as normal text. In one demo with Claude Fable 5, a session that would have cost $42.21 dropped to $6.06.

What’s behind it: AI models with vision don’t “read” images the way they read text — they process them through a separate vision encoder. It turns out that encoder is efficient enough at recognizing rendered text that the model can still work with it. Deepseek explored a similar idea earlier with an OCR system that compresses documents by up to a factor of ten. If the trick catches on, AI companies could simply raise image processing prices — the loophole exists at their discretion.

A fair caveat: the method is lossy. Exact strings like hashes can come back garbled, and processing is slower because of the vision encoder detour. Chong’s own benchmarks show Fable 5 handles the rendered images with 100 percent accuracy on test tasks, while older models like Opus 4.7 and 4.8 misread about 7 percent — which is why they’re off by default.

What this means for you: If you’re just chatting with Claude on a normal subscription, nothing changes — this is for people paying per token through the API or Claude Code. If that’s you and your bills sting, pxpipe is free to try and the savings are documented in the repo. Just don’t point it at anything where a garbled character would hurt, and treat it as a clever experiment rather than infrastructure: it could stop working whenever pricing changes.

Sources

• pxpipe on GitHub
• pxpipe benchmarks and findings
• The Decoder