# pxpipe: Cut Your Fable 5 Bill by 70% With One Proxy

> A local Claude Code proxy that renders bulky text as images so Fable 5 bills pixel tokens instead of 92k-character tool dumps.

- Author: AI Kai
- Published: 2026-07-04
- Updated: 2026-08-24
- Topic: Claude Code
- Tags: claude-code, tokens, proxy, cost
- Canonical: https://hqman.me/blog/pxpipe-token-hack/

## The Problem

Claude Code is expensive. Every request ships your system prompt, tool definitions, and conversation history as text tokens. A single dense tool result burns 92,000 tokens.

## The Insight

Claude charges image tokens by pixel dimensions, not by how much text is inside.

A 1928x1928 PNG costs 4,761 tokens regardless of content. That same image holds ~92,000 characters of dense code. As plain text, those characters cost ~92,000 tokens.

## What pxpipe Does

A local proxy between Claude Code and the Anthropic API. It intercepts requests, renders bulky text into PNGs, then forwards the lighter request.

Gets compressed:

- System prompt + tool docs (repeated every request)
- Large tool_result bodies (file reads, command output above ~6k chars)
- Older conversation history

Stays untouched:

- Your recent messages
- Model output
- Sparse prose (a profitability gate skips when the math doesn't work)

## The Numbers

- Dense tool result: 92,000 tokens to 4,761 tokens
- End-to-end bill (13,709 requests): $100 to ~$41
- SWE-bench Lite (n=10): 10/10 both arms, cost from $53.61 to $27.27
- SWE-bench Pro (n=19): same completion rate, 60% smaller requests
- Gist recall (decisions, values, paths): 98/98 both arms, zero loss

Fable 5 reads image-rendered content at 100/100 on arithmetic tasks.

## The Catch

Lossy. Exact strings (hex IDs, hashes, secrets) are not byte-safe from images.

- Fable 5 verbatim 12-char hex recall: 13/15
- Opus: 0/15

Misreads don't throw errors. The model gives a confident wrong answer. For coding tasks this works fine because the agent re-reads files before editing. For exact identifiers, route to a non-allowlisted model as escape hatch.

## Model Support

- Fable 5: Default ON, 100/100 read accuracy
- GPT 5.6: Default ON
- Opus 4.7/4.8: Opt-in (~7% error rate)
- GPT 5.5: Not recommended

## Worth Noting

pxpipe logs both compressed and uncompressed token counts per request (`.pxpipe/events.jsonl` in your home directory), so you can verify savings yourself. One risk: this exploits current pricing. Anthropic could adjust image token pricing anytime.

Credit: @digimaga

GitHub: https://github.com/teamchong/pxpipe