A tool cuts your model bill 60% by turning text into pictures. The thread says it's arbitrage on a pricing bug, not compression
pxpipe renders your prompts into dense PNGs before sending them, because vision tokens are billed by pixels and text by characters. It really does slash the bill. It also silently confabulates when it hits a password — and, the thread notes, Gemini already gives you the same tokens for free.
pxpipe reflows text into ~1928px columns (~92,000 chars/page ≈ 4,761 vision tokens) and only converts where the math beats ~19 chars/token, claiming a 59–70% lower end-to-end bill. It's explicitly lossy: imaging a 12-char hex string recalled correctly only 13 of 15 times.
Commenters call it a token-accounting loophole, not efficiency. Gemini already OCRs PDFs and doesn't charge for the extracted text; a developer who tried the same trick last year got MORE completion tokens and a bigger bill.
What pxpipe does, and where it breaks
pxpipe is a local proxy that intercepts your API calls and converts token-dense text (system prompts, tool docs, code, logs) into PNG images before they're sent, exploiting that images are priced by pixel dimensions while text is priced per character. It keeps recent turns as text and only images the parts where the math favors it (roughly 19+ chars per token). The claim: 59–70% off the end-to-end bill; one trace fell from $100 to about $41 across 13,709 requests.
It reports strong task benchmarks — arithmetic and SWE-bench Lite holding at large token cuts — with one loud exception it discloses honestly:
| Task | Result imaged | Note |
|---|---|---|
| Arithmetic | 100% correct | at −38% tokens |
| State tracking | 18/18 | — |
| SWE-bench Lite | 10/10 | −65% request size |
| Verbatim 12-char hex recall | 13/15 | "silent confabulations, not errors" |
That last row is the warning label: never image a hash, password, or byte-exact value.
"It's a loophole in token accounting"
The thread liked the cleverness and distrusted the premise. The most deflating comparison came from how another vendor already prices this:
"In Gemini… if you look at how they process PDFs, they do an OCR and then feed the text + image to the model, without charging you for the text tokens… so this hack is probably more of a loophole in token accounting that might get closed" — a commenter describing Gemini's PDF ingestion
Someone had already run the experiment — with the opposite outcome:
"I tried the same thing last year (with openai models), back then it worked to reduce prompt tokens, but you needed way more completion tokens, ultimately more expensive (and slower)" — a developer who replicated the technique a year earlier
The cleanest frame was the cobra-farm one:
"It works, and it's clever, but… it's clearly a workaround for a pricing failure. Much like the bounty on poisonous snakes leading to people taking up snake-breeding, this just exploits and promotes waste… blame falls on Anthropic for the poor pricing" — a commenter reading it as perverse-incentive arbitrage
The implication for anyone tempted to deploy it: you're not buying a compression breakthrough, you're renting a mispriced billing quirk. The day a provider prices vision the way Gemini already prices PDF text, the 60% saving vanishes, and so does any reason to have accepted the confabulation risk.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Ubicloud says the Linux OOM killer can drop every Postgres connection at once. Its own co-founder says the post oversells it
When Linux kills one Postgres backend to reclaim memory, shared state can go inconsistent — so the postmaster kills all of them and every connection dies. Ubicloud's fix is one sysctl. Then Ubicloud's co-founder showed up in the thread to say the title came on too strong.