Save up to 60% on tokens with hyped Fable 5: developers found a way to game the system and pay Anthropic far less

The tool is a local proxy that intercepts requests to Claude Code and converts static, heavy content — system prompts, tool documentation, and older chat history — into dense PNG pages.

The trick exploits how vision and text tokens are priced differently: reading an image costs a fixed number of tokens based on its pixel dimensions, not on how many characters of text it contains. Because of this, packing large volumes of text into a single densely written image can end up far cheaper than sending the same content as plain text.

In the developer’s own demo, a task that normally cost $42 in text tokens was completed for about $6 once routed through pxpipe — roughly seven times cheaper. Try here.

Claude Fable 5 reads the rendered text reasonably well, but the method is not lossless. On tests involving exact strings, such as 12-character hexadecimal codes buried in imaged content, the model got them right most of the time but not perfectly, and its misreads come with no warning — it simply produces a confident wrong answer instead of flagging uncertainty.

Developers are advised to keep exact IDs, hashes, and secrets in plain text rather than imaging them. Silent misreads are worse than a visible error, since a wrong hex code that looks confident can slip through review far more easily than an obvious failure.

For teams running high-volume Claude Code sessions where system prompts and documentation dominate the token bill, this is worth testing on non-critical context first, keeping anything that needs to be exact — credentials, hashes, precise identifiers — in plain text regardless of the savings elsewhere.