We Retracted the “70-80% Token Savings” Claim

The first draft of the Snapdown page had a line: “70-80% token savings”. In the second draft, the number became 60%. Neither was ever measured. What happened is what usually happens: someone estimated, the estimate got written down, and the writing became fact simply because it was written.

We retracted both before anyone called us on it.

Why the Number Could Not Stand

Token use depends on things that all vary: the model, the agent product around it, how it counts image input, and how much of the conversation it carries over each turn. A single percentage for all of that is not a measurement. It is an average of something that was never counted.

A number that ships must be verifiable by the reader on their own machine. If it can't be, it isn't a number: it's a promise.

We had already written that rule into our writing rules before this case came up. What was new was using it against our own sentence.

What Snapdown Does Instead

What Snapdown controls is the image itself. When it saves a capture, it scales the image down so its longest edge fits the chosen preset, then encodes it again at the preset's quality. Both steps run on your computer. The first step removes pixels; the second makes each remaining pixel cost fewer bytes.

The pixel part is plain arithmetic you can redo. A full-screen 4K capture is 3840 by 2160 pixels, about 8.3 million. Here is what each preset keeps of that frame:

We give no byte figure: the presets shrink resolution and file size, and each preset and its quality setting is listed in the Snapdown docs.

Token Savings Are Experimental

Vision models count image input by its pixel size, so a smaller image should take less of an agent's context. How much less depends on the model and the product, and we have not measured it. That is why we call token savings experimental and give no percentage.

If you compare the same finding at two presets in your own agent, we would like to hear what you saw: support@wiradelta.com.

What Changed on the Product Page

What We Learned

An unmeasured number doesn't stay in one place. It gets copied. By the time we traced it, it was already in three places with three different values, and no one had noticed, because no one had ever compared all three.

Now a number we publish is one the reader can check: the pixel counts above follow from the preset table, and anyone can recompute them.