Breaking

pxpipe cuts Claude Fable 5 costs up to 70% by rendering code context as images for OCR

July 3, 2026 at 20:21 EDT

Software Dev & Coding
AI Agents
Foundation Models

A developer tool called "pxpipe" is drawing attention in early July 2026 for cutting Anthropic's Claude Fable 5 API costs by up to roughly 70% by converting the code context sent to the model into PNG images and having it read them back via OCR. It runs as a local proxy and is published on GitHub.

Open-Source Tool · Claude Fable 5

pxpipe turns text into images to cut Claude Fable 5 costs by up to ~70%

A local proxy renders most context into PNG images and lets the model read them via OCR — compressing token counts. In one demo, a single session fell from $42.21 to $6.06.

59–70%

cost reduction, depending on workload

~3.1×

chars per image token vs ~1 char per text token

token context length exploited by Fable 5

Same session, the cost drawn to scale

One demo session, before vs after pxpipe compression.

$42.21

without pxpipe

$6.06

with pxpipe

≈86% saved in this run — about 7× cheaper for the same answer.

Why the tokens shrink

~48,000 chars of dense text, encoded two ways.

As text tokens~25,000

As image tokens~2,700

How the proxy decides

Old history & system prompts

→

Rendered to PNG, read via OCR

Recent turns, IDs, hashes

→

Kept as exact text

WHAT WORKS

Same answer for ~7× cheaper on dense code
10/10 solved on SWE-bench Lite at half the cost
Quick fix for Max-plan allowance burning too fast

THE CATCH

Lossy: misreads on exact numbers, IDs, hashes
Accuracy drops especially with Opus 4.8
For sparse English text it can cost more

Fable 5 pricing pain point

At $10 input / $50 output per million tokens — roughly double Opus 4.8 — long agent sessions get expensive fast. Combined with prompt caching, input costs can drop by up to 90%, but effects vary — measure in your own environment.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.