How pxpipe Cuts Fable 5 Token Costs by Up to 70%

The author reports a technique named pxpipe that converts the textual context of Anthropic’s Fable 5 model into tightly packed PNG images, then feeds those images to the model’s vision (computer‑use) channel. Because image token cost is based on pixel dimensions rather than character count, the same amount of information can be represented with far fewer tokens.

In a concrete example, a 48 000‑character system prompt and tool documentation would require about 25 000 text tokens, but after rendering to a 1573×1248 image it consumes only roughly 2 700 image tokens. Using the current Fable 5 pricing, the end‑to‑end bill drops by 59 %–70 %.

This thing turns Fable 5’s text context into images to lower token usage.

The tool is a local proxy that intercepts Claude Code requests, rewrites bulky context (system prompts, tool docs, early history, large tool outputs) into compact PNGs, and inserts them back into the request. The model reads the images via the same visual pathway used for screen‑shot OCR, so no separate OCR step is needed.

Benchmark results show a dramatic cost saving: the original text‑only run cost $42.21 and filled 96 % of the context window, while the pxpipe version cost $6.06 and used only 73.5 k/1 M image tokens, completing the same task successfully.

However, the approach is lossy. When the same experiment is run on the Opus‑4.8 model, the image‑based context works for normal text but fails on word‑frequency statistics, indicating that precise string extraction can break. Errors may not be flagged; the model can hallucinate confident but wrong answers.

pxpipe therefore keeps exact identifiers, hashes, keys, numbers, and recent dialogue as text, compressing only the long, dense sections.

To try it yourself, the author provides a two‑line startup command:

npx pxpipe-proxy # proxy on 127.0.0.1:47821
ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude # point Claude Code at it

After launching, a local dashboard (http://127.0.0.1:47821/) shows token savings per conversion, a kill‑switch, and the active model tag. Note that pxpipe only compresses the request side; model outputs are returned unchanged.

Related research shows this idea is not new. Google’s 2022 CLIPPO paper rendered text as RGB images so a single Vision Transformer could process both visual and textual inputs, achieving near‑CLIP performance with only a 1‑2 % drop. DeepSeek‑OCR also demonstrates visual compression of long contexts, achieving ~97 % decoding accuracy at <10× compression and ~60 % at ~20×.

These works together explain why converting text to images can be an effective cost‑saving strategy for vision‑enabled LLMs like Fable 5, while also highlighting the trade‑off: loss of exact string fidelity and dependence on the model’s visual reading capability.