After Fable 5’s return, many people will definitely want to try just how powerful it is, but anyone who’s used it knows that it consumes a lot of tokens—it’s the kind of model you need to be especially careful with.
Recently, a tool called “pxpipe” has gained significant attention on GitHub. When paired with Claude Code and Fable 5, it claims to reduce overall request costs by approximately 59% to 70%. Rather than trimming content, it renders large amounts of text into high-density PNG images, which are then read by Fable 5’s visual reading capability, preserving the original context information at a lower image token cost.
pxpipe Embeds Long Text in PNG to Save Claude Code Tokens: Fable 5 Benchmark Shows Up to 70% Savings
pxpipe is an open-source proxy tool that runs locally, helping reduce input token costs for AI development tools like Claude Code.
Put simply, once you launch pxpipe, the requests that Claude Code would normally send directly to the model API are first routed through pxpipe for inspection. If there’s a large amount of text content suitable for compression, pxpipe converts it into high-density PNG images, then passes them to models that support visual reading capabilities.
約 48,000 個字元的系統提示詞加工具文件，如果直接將文字送進模型，大約會消耗 25,000 個文字 token。但如果先渲染成一張高密度 PNG 圖片，約只需要 2,700 個影像 token。
However, pxpipe doesn’t convert everything into images—it only selects and compresses the long, token-heavy background information that doesn’t necessarily need to be read word-for-word, such as: tool output results, file contents, terminal logs, earlier conversation history, system prompts, and tool documentation, etc. Therefore, the user’s latest input, recent conversation turns, model responses, precise numbers, and other content that absolutely cannot be misread will be preserved as text whenever possible.
Developers say pxpipe can reduce overall billing costs by approximately 59% to 70% in real-world usage.
Out of 13,709 requests, the original cost of approximately $100 dropped to approximately $41 after using pxpipe, representing approximately 59% savings. Another set of tracking data for 8,904 compression requests measured approximately 70% cost reduction.
The difference is even more apparent in the demo video provided on the GitHub page. For the same Fable 5 session, without pxpipe the cost is $42.21, and the context window is almost full. After using pxpipe, the cost drops to $6.06, and the context window still has more available space:
Below is the official benchmark. pxpipe’s performance on Fable 5 is relatively stable. In tests like novel arithmetic, gist recall, and state tracking, the image version and text version results are nearly identical, with some tests reducing tokens by 38%. However, in tests requiring character-by-character precision, such as 12-character hex strings, the Fable 5 image version only scored 13/15, and Opus 4.8 scored 0/15. This also shows that pxpipe is suitable for compressing large amounts of background context but is not suitable for handling critical strings that must be completely accurate:
Using pxpipe is also quite simple. First, run npx pxpipe-proxy in the terminal to start the local proxy, then point Claude Code’s ANTHROPIC_BASE_URL to http://127.0.0.1:47821, and start Claude normally. After that, pxpipe will automatically determine which content is suitable for image-based compression, and you can also go to the Dashboard to check the token savings status.
One important thing to note here is that pxpipe is not enabled by default for all models. Currently, the models with default support are claude-fable-5 and gpt-5.6, because these two models are better at stably reading contextual content that has been converted to images. As for Opus 4.7 and Opus 4.8, official testing found approximately a 7% misreading rate when reading image content; GPT 5.5 also performs poorly with image-based contexts, so these models are not automatically enabled and require users to manually add them to the supported list.
However, for regular users, it’s still not advisable to treat it as an essential tool at this point, since image-based context is less reliable than plain text. Additionally, this approach is susceptible to changes in the model’s image comprehension capabilities and safety mechanisms, so it’s currently better suited as an experimental tool rather than a stable cost-saving plugin.
Source: KOCPC Chinese