60% Fable cost cut by converting code to images and having the model OCR it(github.com)
284 points by dimitropoulos 21 hours ago | 92 comments
tl;dr: pxpipe is a local proxy that rewrites bulky Claude Code request context (system prompts, tool docs, older history) into PNG images, exploiting the fact that dense text packs ~3.1 chars per image-token vs ~1 char per text-token, cutting end-to-end costs ~59-70% on Fable 5. The tradeoff is lossy verbatim recall — exact strings like hashes and IDs can silently confabulate, so it's gated to token-dense content and specific models (Fable 5, GPT 5.6), with recent turns and byte-exact data kept as text. SWE-bench results show near-parity with 60-65% token reduction.
HN Discussion:
  • ~This is a pricing loophole that providers will likely close once they optimize token accounting
  • Tried this approach before and found it increased completion tokens making it more expensive overall
  • This is wasteful pricing arbitrage that promotes digital junk and exploits a pricing failure
  • Using images for agent context is genuinely effective and leads to better workflows
  • Related prior work and techniques exist (DeepSeek paper, other tools) validating the approach