Text Tokens Are a Scam. Here’s How One Developer Cut AI Costs 60% by Sending Images Instead.

You’re paying for text tokens. Someone else is paying for image tokens. Guess who’s getting robbed?

A developer using Anthropic’s Claude API just slashed their costs by 60% — not by optimizing prompts, not by switching models, but by converting their text into images and letting the model OCR it back. The project, called pxpipe, is live on GitHub right now. And it exposes something the AI industry doesn’t want you to think too hard about.

Here’s the trick: when you send text to an LLM, you’re charged per token. But when you send an image, the pricing model treats it differently — often cheaper for the amount of information it carries. So pxpipe takes your text, renders it as an image, sends that image to Claude, and lets the model’s vision capabilities do the heavy lifting. The result? Same information in, same intelligence out, fraction of the cost.

The pricing models we’re building billion-dollar businesses on are held together by accounting duct tape.

This isn’t some fringe experiment either. There’s a DeepSeek whitepaper on the same technique. Other developers tried it last year with OpenAI’s models. People have been poking at this hole for a while — pxpipe just made it dead simple to use.

But here’s where it gets uncomfortable. When you convert text to an image, you’re actually burning MORE compute. The model has to run vision processing, OCR the image, extract the text, and THEN process it. You’re using more silicon, more electricity, more GPU time — all to pay less money. It’s the AI equivalent of driving 20 miles out of your way to save 3 cents on gas.

That paradox is the whole point. The pricing doesn’t reflect the cost. It reflects what providers think they can charge before you complain.

One commenter on the project nailed it: this is a pricing hack that burns resources, and when the loophole gets closed, the price of OCR-style processing will have to rise. They’re right. Anthropic, OpenAI, Google — none of them are charities. The moment enough people exploit this, they’ll patch it. And when they do, they won’t just close the loophole. They’ll likely raise prices across the board to cover the “resource abuse” they just discovered.

Every exploit is a stress test, and every stress test ends with the vendor rewriting the rules.

There’s also a darker possibility lurking here. A commenter pointed out that Gemini’s backend already does something similar when processing PDFs — it runs OCR internally and feeds text plus image to the model without charging for the text tokens. So maybe Claude is doing the same thing. Maybe pxpipe isn’t exploiting a bug — it’s exploiting a feature that was never supposed to be user-facing. The pricing inconsistency might be an accounting artifact of how these companies internally handle multimodal inputs.

Which means the real question isn’t “how long until they close this loophole?” The real question is: if the true cost of processing is higher than what they’re charging for image tokens, how long until the entire pricing structure collapses and gets rebuilt from scratch?

For developers and businesses building on LLMs, this is a double-edged sword. Yes, you can use pxpipe right now and cut your API bill dramatically. It works. People have verified it. But you’re building on a foundation of sand. The day Anthropic adjusts their token accounting — and they will — your architecture breaks, your costs spike, and you’re left explaining to your CFO why the bill tripled overnight.

The smart play isn’t to exploit this. The smart play is to understand what it reveals. The AI pricing landscape is arbitrary, inconsistent, and ripe for disruption. The companies charging you per token are guessing at what you’ll tolerate. The moment enough people find the cracks, the whole edifice shifts.

Don’t optimize for the loophole. Optimize for the day the loophole disappears — because that’s when you’ll find out what everything actually costs.