A developer built an open-source proxy called pxpipe that can cut Claude API bills by turning bulky text context into PNG images. The idea works because image tokens are priced by pixel size, not by how much readable text is inside the image. According to the project, a dense 1928 by 1928 image can hold around 92,000 characters while costing about 4,761 vision tokens. That creates a strange arbitrage. In one production snapshot, pxpipe says a bill dropped from $100 to about $41 by converting older context, tool docs, and system prompts into images before sending them to The catch is accuracy. The compression is lossy, and the failure mode can be silent. Fable 5 recalled exact 12-character hex strings correctly in 13 of 15 tests, while Opus failed all 15 in the same dense render test. This makes pxpipe useful for bulky context, but risky for IDs, hashes, secrets, or anything that needs exact recall. Would you trust this for real coding work? Source: GitHub, teamchong/pxpipe