pxpipe is an open-source local proxy that converts the bulky text blocks in Claude Code requests (system prompt, tool docs, old conversation history) into compressed PNG images and sends those to the API instead. Because an image's token cost is fixed by its pixel size, it significantly lowers input token cost on text-heavy workloads.

Does the image-context method really save 60%?

It depends on the workload. In the project's own measurements, end-to-end billing drops 59-70% on token-heavy content like code and logs; but on sparse prose the method loses money and the tool automatically skips it. So the 60% figure belongs to specific workloads, not everyone.

What is the risk of this method?

There are two core risks: first, the method is lossy — exact critical values like IDs, hashes and amounts in the imaged content can be silently misread. Second, it is a pricing arbitrage; when the model provider changes its pricing model, the gain can vanish overnight. Instead of tying critical workflows to this, you should prioritize durable methods like prompt caching and the batch API.

The Image Trick That Cuts AI Token Costs by 60%: Does pxpipe Actually Work?

A "cost hack" spread quickly through AI automation circles recently: if you send your long prompts not as text but as an image written in a very small font, Claude reads the same content but you're billed far fewer tokens. We examined the open-source tool behind this claim (pxpipe — GitHub) and the data it published. The verdict: the method is real, the numbers are largely correct — but there are serious details you need to know before turning it into a business decision.

How Does the Method Work?

In models like Claude, text is split into tokens based on its content and charged per token. Images are different: an image's cost is fixed roughly by its pixel area — whether it contains 100 words or 15,000, the price is the same.

pxpipe exploits exactly this asymmetry. It intercepts the request going from Claude Code to the API with a local proxy; it converts bulky blocks like the system prompt and old conversation history into small-font PNG pages and forwards the request that way. The measurement is striking: a single 1928×1928-pixel image costs about 4,761 vision tokens but fits ~92,000 characters — the same content sent as text would run roughly 25,000+ tokens.

A critical nuance: the model isn't doing "OCR" in the classic sense, it interprets the image directly. The newest models (especially Fable 5) read text at this density surprisingly well — in the project's benchmarks, on brand-new numerical problems that can't be memorized, the image and text setups give the same accuracy (100/100).

What Do the Numbers Say?

The "30% savings" videos circulating on social media are actually the conservative end of the story. The end-to-end bill drop the project measured on its own production logs — that is, the full account including untouched small requests, cache writes/reads and output tokens — is between 59% and 70%. The gain is highest on "token-heavy" content like code, log files and JSON; on sparse prose the method mathematically loses money, and the tool automatically skips imaging in that case.

So the correct reading is this: 60%+ savings isn't a figure everyone will see, but a ceiling that belongs to workloads carrying long technical context. In agency practice that translates to systems running constant queries over knowledge bases of hundreds of thousands of tokens — assistants analyzing customer data, bots scanning large document archives, long-running code agents.

The Honest Part: Why We Don't Recommend It to Everyone

The project itself doesn't hide this — its documentation has a section literally called "The honest part." Let us relay it with the same honesty:

The method is lossy.Exact values in the imaged content (order number, IBAN, hash, amount) can be misread. In the best case, even on Fable 5, 2 of every 15 twelve-character codes in dense images come out wrong. Worse: the model doesn't say "I couldn't read it," it confidently makes up the wrong value. For flows carrying billing, contract or identity data, this is an unacceptable risk.
Model dependent.Opus 4.8 practically can't read the same images at all (0 of 15 in the same test). The gain currently leans on a single model's image-reading ability.
This is a pricing arbitrage.The source of the gain isn't a technological efficiency but a gap in image pricing. The day the provider makes image pricing sensitive to text density, this gain drops to zero. Building a critical business process on a gap of uncertain lifespan is gambling, not engineering.

Do These First: Saving Without the Risk

In our experience, most businesses with ballooning AI costs aren't using the official methods sitting on the table before they reach for hacks like this. The order should be:

Prompt caching:Cache the repeating system prompt and knowledge base — a token read from cache costs a tenth of the normal price. In most chatbot and assistant scenarios that alone means 50%+ savings.
Batch API:Queue work that doesn't need an instant response (reporting, content generation, data labeling) into a batch — it drops to half the list price.
Right model choice:Not every task needs the biggest model. A tiered architecture using a small model for classification and simple replies and a large model for complex reasoning lowers the bill permanently.
Context hygiene:Instead of growing the conversation history without limit, carry it forward summarized and prune needless tool output — the cheapest token is the one never sent.

After applying these four steps, if you're still running queries of hundreds of thousands of tokens and there are no exact critical values in the content, pxpipe-style image compression may be an extra layer worth trying — not the main strategy.

An Agency's Verdict

pxpipe is good proof of the truth that "AI costs aren't fixed, they're a design decision." The monthly bill difference between two automations doing the same job can be 5–10x, and that difference almost always comes from architecture: what gets cached, what gets batched, which model is used where. The image-context trick is the newest and most fragile tool in that toolbox — know it, try it if you must, but don't rest your business's backbone on it. As we said in our post on what Fable 5 brings: the model side is changing fast; the durable advantage is in well-built infrastructure.