pxpipe cuts Claude Code and Fable 5 token costs up to 70%

pxpipe converts long text inputs for Claude Code and other models into compact PNGs to reduce token costs, with reported savings averaging 59 to 70 percent and a demo session that fell from $42.21 to $6.06.

Developer Steven Chong published the open-source proxy and documented benchmarks on GitHub on Jul 4, 2026. The tool intercepts requests and renders bulky, static parts of prompts and documentation as images, while recent messages and model outputs continue as text.

How does pxpipe cut token costs?

pxpipe packs dense, static content into PNG pages because Anthropic charges images by pixel dimensions rather than by character or token count. That pricing quirk lets a single image token represent many characters: pxpipe can achieve about 3.1 characters per image token.

In practice pxpipe runs as a local proxy that replaces system prompts, tool documentation, and older chat history with rendered PNGs. The source gives an example where roughly 48,000 characters of system prompt and tool documentation would cost about 25,000 tokens as text, but the same content becomes roughly 2,700 image tokens when rendered onto a single dense PNG page. The proxy leaves recent messages and model outputs as plain text so the model still sees the conversational context directly.

What are the accuracy and speed tradeoffs?

The technique is lossy and slower: exact strings such as hashes can be garbled when read from images, and models must run vision encoders instead of reading text directly. That increases latency and introduces reading errors for some models.

Benchmarks in the repository show mixed results across models. Fable 5 reached 100 percent accuracy on math problems that used fresh random numbers the model could not have memorized. By contrast, Opus 4.7 and 4.8 misread about 7 percent of the rendered images, and GPT 5.5 "also does worse with image context." The repository notes that Opus and GPT image-based modes are off by default and must be enabled manually. pxpipe ships with support for Claude Fable 5 and GPT 5.6 out of the box, and the README links to further evaluations.

Why it matters

pxpipe exposes a direct arbitrage between text and image pricing that can meaningfully lower per-session bills for workflows with large static context like system prompts and tool docs. The tool shrinks what would be tens of thousands of text tokens into a few thousand image tokens, which delivered a reported per-session cost drop in one Fable 5 demo from $42.21 to $6.06. If the trick gains traction, platforms could respond by changing image pricing, removing the savings vector.

The tradeoff matrix matters for teams that rely on exact string fidelity or low latency. Where exact hashes, precise identifiers, or fast turnaround matter, the lossy reads and vision-encoder overhead will negate the cost benefit. Where large static context dominates cost and exact character-for-character fidelity is not required, pxpipe can shift the economics.

What to watch

Watch for two signals: whether adoption of pxpipe-style proxies grows among heavy-context users, and whether model providers change image processing prices or token accounting. The GitHub repository documents the benchmarks and evaluations that will clarify which models and tasks keep accuracy while delivering the biggest savings.

Specific source-attributed data points in the repository and reporting include the average reported savings of 59 to 70 percent, the Fable 5 demo drop from $42.21 to $6.06, the example compression from about 25,000 text tokens to roughly 2,700 image tokens for 48,000 characters, the roughly 3.1 characters per image token figure, Fable 5 hitting 100 percent accuracy on certain math benchmarks, and Opus 4.7/4.8 misreading about 7 percent of rendered images.

For teams weighing the tradeoffs, the next practical step is to test pxpipe on their own prompts and tool docs and compare throughput, latency, and string fidelity against the token-cost savings shown in the published benchmarks.