Open-source tool cuts Claude, GPT token costs 59–70% by hiding text in images

What happened

Developer Steven Chong released pxpipe, an open-source proxy that converts long text inputs—system prompts, code, chat history—into compressed PNG images before sending them to Claude Code and other AI models. The trick exploits Anthropic's pricing: text costs roughly one token per character, but images cost a fixed number of tokens regardless of content density, allowing about 3.1 characters per image token. In one Fable 5 demo, session costs fell from $42.21 to $6.06.

Why it matters

For businesses and developers using Claude Code or GPT 5.6 at scale, token costs directly affect operational spending. This technique could reduce invoices by 59 to 70 percent on many workflows. However, the approach is lossy—exact strings like hashes can become garbled when read from images, and processing is slower because the model must run images through a vision encoder instead of reading text directly. Fable 5 hits 100 percent accuracy on math benchmarks, but Opus 4.7 and 4.8 misread about 7 percent of rendered images, and GPT 5.5 also performs worse with image context.

What to watch

pxpipe supports Claude Fable 5 and GPT 5.6 by default; Opus 4.7, 4.8, and GPT 5.5 can be enabled manually. If this technique catches on widely, AI companies may respond by raising image processing prices. The approach is not new—Deepseek built a similar OCR system that compresses text documents by up to a factor of ten while retaining 97 percent of information.