pxpipe is built around a simple but unusual idea: some long text sent to an AI model may be cheaper when it is shown as an image instead of passed in as ordinary text. The open-source tool applies that idea to Claude Code by converting large, static context into densely packed PNGs.
The result is a practical attempt to reduce token spending in long coding sessions. It does not make the model read less information. Instead, it changes the form of that information, taking advantage of the way image inputs are priced.
Why PNG context can lower token costs
The cost difference comes from how Anthropic prices text and images. Text is described as costing roughly one token per character. Images, by contrast, are priced by pixel dimensions, regardless of how much text is packed into the image.
That creates an opening for dense material such as code, JSON, system prompts, tool documentation, and older chat history. When this kind of content is rendered into an image, the source says it can fit about 3.1 characters into every image token.
pxpipe turns that pricing gap into a workflow. Rather than sending all context as normal text, it acts as a local proxy and intercepts requests to Claude Code. It then decides which parts are suitable for image conversion and which parts should stay as text.
The tool focuses on bulky, static content. Recent messages and model outputs continue to pass through as normal text, which matters because those parts are active parts of the conversation. Older or reference-like material can be compressed visually without changing the user-facing flow.
What the model sees
The source describes an example where around 48,000 characters of system prompt and tool documentation are rendered onto a single densely packed PNG page. Sent as text, that material would cost about 25,000 tokens. As an image, it is roughly 2,700.
That example explains why the approach is drawing attention. Long-running AI coding sessions can accumulate a lot of context, and much of that context is important but not always changing. pxpipe tries to preserve that context while reducing the token bill attached to carrying it forward.
According to developer Steven Chong, total savings average 59 to 70 percent. In one Fable 5 demo, session costs dropped from $42.21 to $6.06. Those numbers are the core promise of the tool: not a small optimization, but a major reduction in spending under the right conditions.
The method also creates a possible pricing tension. The source notes that if this exotic trick catches on, AI companies could respond by raising image processing prices. In other words, the economics work because of a current pricing difference, and that difference may not be permanent.
The tradeoff is accuracy
pxpipe is not presented as a lossless compression system. The source is clear that the approach has downsides, and the most important one is reliability. When text becomes an image, the model has to recover the information visually.
That matters for exact strings. Hashes and similar precise values can come back garbled when read from images. For software work, that limitation is important because a single wrong character can change the meaning of a command, identifier, checksum, or data value.
The practical takeaway is that pxpipe fits better for bulky context that gives the model background, structure, or documentation. It is riskier for material where exact reproduction is the main requirement. The tool’s design partly reflects that distinction by keeping recent messages and model outputs as normal text.
Benchmarks and evaluations are documented in the repository. Fable 5 hits 100 percent accuracy in benchmarks on math problems with fresh random numbers the model cannot have memorized. According to Chong, Opus 4.7 and 4.8 misread about 7 percent of the rendered images, and GPT 5.5 also does worse with image context.
Those model differences shape the defaults. By default, pxpipe supports Claude Fable 5 and GPT 5.6. Opus 4.7, Opus 4.8, and GPT 5.5 are off by default and can only be enabled manually.
The tradeoff is also speed
Cost is not the only variable. Processing is slower because the model has to run the rendered PNGs through a vision encoder instead of reading text directly. That adds work to the request even when the token accounting looks better.
For users, this means pxpipe is a trade between money, latency, and precision. A session with large static context may benefit from the savings. A session that depends on fast iteration or exact string handling may feel the cost in other ways.
The source also places pxpipe in a broader pattern. Feeding text to AI models as compressed images is not a new idea. Deepseek built an OCR system that processes text documents as images and, according to its technical paper, compresses them by up to a factor of ten while keeping 97 percent of the information.
pxpipe applies a related idea to the economics of coding agents and long context. Its value is clearest where the input contains a lot of repeated, static, or reference material. Its limits are clearest where the model must read every character perfectly.
What pxpipe changes
The most interesting part of pxpipe is not just that it reduces token costs. It shows how tool builders are beginning to optimize around the pricing rules of multimodal models, not only around model capability.
That makes the technique both useful and fragile. It can produce large savings under current conditions, but it depends on image pricing, model vision accuracy, and the kind of context being sent. For Claude Code and Fable 5 users handling long sessions, pxpipe offers a concrete experiment in lowering costs while keeping the essential context available.