Developer Steven Chong has released pxpipe, an open-source tool that converts text context sent to Claude into PNG images to sharply reduce input token charges, reportedly cutting bills by up to 70% on production workloads.
Open-Source Tool · pxpipe
Turning text into images to cut Claude bills by up to 70%
A local proxy for Claude Code renders bulky, repeatedly-resent context — system prompts, tool definitions, chat history — into dense PNGs and has Claude read them as vision, exploiting the pricing gap between text and image tokens.
70%
peak reduction in coding-session bills
65%
smaller requests on SWE-bench Lite
1,200+
GitHub stars by July
One real trace: a $100 bill over 13,709 requests
Compression applies only to input tokens — output is untouched, so savings vary by workload.
How it works
Runs locally at 127.0.0.1:47821 — no code changes, just set ANTHROPIC_BASE_URL.
1 · CAPTURE
Bulky resent text: prompts, tool defs, history
→
2 · RENDER
Pack text into a dense 1928×1928 PNG
→
3 · READ AS VISION
Claude parses image in 28×28px visual tokens
Why it saves: characters packed per token
A 1928×1928 image holds ~92,000 characters in ~4,761 visual tokens.
~1 char
plain text
~3.1 chars
image token
The upside
Sharp drop in coding-session bills
SWE-bench Lite: 10/10 passed, both arms
Free, MIT-licensed, no code changes
Dashboard shows real-time savings + kill switch
The catch — it's lossy
Dense hex recall: Fable 5 got 13/15, Opus 0/15
Silent confabulation is possible
IDs, hashes, secrets must stay as text
Opus opt-in; short prose can backfire
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…