Updated
Updated · THE DECODER · Jul 4
pxpipe Cuts Claude Code, Fable 5 Costs Up to 70% by Turning 48,000 Characters Into PNGs
Updated
Updated · THE DECODER · Jul 4

pxpipe Cuts Claude Code, Fable 5 Costs Up to 70% by Turning 48,000 Characters Into PNGs

3 articles · Updated · THE DECODER · Jul 4

Summary

  • Steven Chong’s open-source pxpipe routes Claude Code prompts through a local proxy that converts bulky text into PNGs, cutting total session token costs by 59% to 70%; one Fable 5 run fell to $6.06 from $42.21.
  • Anthropic’s image pricing drives the savings: text costs about one token per character, while images are billed by pixel size, letting dense rendered content pack roughly 3.1 characters per image token.
  • pxpipe targets static prompt material such as system prompts, tool docs and older chat history, while recent messages and model outputs stay as plain text; about 48,000 characters can shrink from roughly 25,000 tokens to 2,700.
  • Accuracy and speed are the tradeoff: Fable 5 scored 100% on benchmarked math tasks, but Opus 4.7 and 4.8 misread about 7% of rendered images and GPT 5.5 also performed worse, so those models are disabled by default.
  • The approach is not new—DeepSeek has reported image-based text compression of up to 10x with 97% information retention—but wider adoption could push AI providers to raise image-processing prices.

Insights

Will AI giants patch the pricing 'loophole' that makes these tools possible, or is this a new normal?
Could a simple image hide the next major AI security breach, bypassing all text-based safety filters?
As AI costs soar, is trading perfect accuracy for massive savings the future of prompt engineering?