A trick made the rounds recently: cut a model’s token bill roughly in half by rendering bulky text, the system prompt, tool docs, history, as an image, because images are billed by dimensions rather than character count. It is a genuinely clever exploit, and it is a symptom of something more useful than the hack itself: in production, your AI cost is dominated by context, and context is a discipline you control.

Token pricing is more arbitrary than it looks

A provider’s real cost is the fixed expense of running the hardware. They bill you per token, but a busy data center is far cheaper per token than an idle one, so the price you see is a business decision layered on top of a fixed cost, not a direct passthrough. That is exactly why a hack like billing text as an image can work at all: the meter is a proxy, not physics. You should not build your operations on exploiting the meter, but you should absolutely notice what the meter is measuring, because it is measuring context.

Where the money actually goes

Every call re-sends context: the instructions, the tools, the relevant history, the retrieved documents. In a real workflow that dwarfs the model’s short answer, and it is charged on every single turn. So the team that stuffs a giant instruction file and the entire codebase into every request pays for that bloat continuously, and, worse, gets a duller model, because more noise in the window means less attention on the task. The expensive setup and the low-quality setup are usually the same setup.

The discipline that pays twice

Managing context is the rare optimization that cuts cost and improves quality at once. Send the model what the task needs and not a byte more: a lean, current instruction set rather than an exhaustive stale one, the specific files that matter rather than the whole repository, retrieved snippets rather than entire documents. This is the same argument as keeping context lean and handing an agent a map instead of a haystack: the shape of the context, not the size, is what matters. Do it well and the bill drops and the answers sharpen.

So before anyone reaches for a billing exploit, look at what you are sending. Most surprising AI invoices are not a pricing problem, they are a context problem wearing a pricing costume, and the fix makes the system better on the way to making it cheaper. That discipline is baked into how we run our solutions, and it compounds nicely as the underlying prices keep falling.

Prompted by the video “Save 50% Tokens With One Simple Trick” (2026). Credit to the creator; the lesson drawn here is ours.