Rendering text as PNGs to save on Claude Code tokens is clever and dumb at the same time

There’s a new open-source proxy called pxpipe that renders Claude Code’s system prompt, tool docs, and older conversation history into PNG images before the request hits the API. People are reporting 60–70 % cuts on Fable 5 and Claude Code usage—just by sending dense non-executable text as a rasterized image instead of raw tokens.

The trick exploits a pricing anomaly in Anthropic’s token accounting. Image inputs cost less per byte of information than an equivalent wall of text, because the model sees a compressed visual representation and Anthropic bills by image tiles. Put differently: the tokenizer is the bottleneck, and packing text into pixels is a compression hack that games the unit cost. That’s not a new idea (I’ve seen people paste logs as screenshots for years), but pxpipe automates it into a proxy, which is what makes it operationally interesting.

The dumb part is betting that a model reading a compressed PNG won’t hallucinate on anything nuanced. A short instruction like “respond in JSON” is probably fine. A verbose API schema with nested types? I don’t trust it. The proxy doesn’t preserve text selectability or structure, and the model is reconstructing intent from a bitmap. If your agent chain relies on exact tool output specifications, you’re adding a new failure mode to save money on a per-request basis.

What I’m watching next: Anthropic will probably close this loophole—either by adjusting how image inputs are tokenized (treating text-heavy PNGs differently) or by nudging the pricing model. More interesting is whether it pushes people toward models with flatter cost structures, like Gemini’s long-context pricing, where you don’t need to play compression theater. The proxy exists because the pricing is wrong, not because PNGs are a great medium for system prompts.