What a weighted token is
One unit that works across every model, so a balance means the same thing whichever you call. Here is the formula, why it exists, and what it buys.
The whole of it
No rounding rules hidden underneath, no minimum charge, no per-request fee. This is the entire calculation.
WT = (in × W_IN + cached × W_CACHED + out × W_OUT) × M_modelInput tokens count as themselves. Cached input counts for much less, output counts for more, and the whole thing is scaled by the model multiplier. The three weights and the multiplier are published per model on the catalog, and the values used to bill a request are the ones captured at the moment it ran, not whatever they are today.
Why not just charge per token
Because a token means different things on different models, and pretending otherwise is how a bill becomes a surprise.
Models cost materially different amounts to serve. A per-token price hides that behind a number that looks comparable and is not: the same thousand tokens can differ by a factor of forty depending on which model you sent them to.
A weighted token makes the conversion explicit instead. Your balance is one number, and the multiplier tells you in advance how far it goes on each model. Choosing a cheaper model makes your allotment go further, visibly, and that is a real benefit rather than a footnote.
The ladder is anchored on glm-5.3 at ×1.0. Everything else is priced relative to it, so re-anchoring the display never changes what a past request cost.
What it buys
Computed from the catalog as it stands right now, not typed into this page.
The walk-in trial is 100,000 weighted tokens. On GLM-5.3 Flash that is roughly 2M input tokens, enough to feel out whether the thing works before any card is involved.
Cached input on glm-5.3-flash is weighted ×0.2, which is why a long agent context costs so much less on its second turn than its first. See prompt caching for what that looked like on a real session.
Every request you make is listed with its own arithmetic on your activity page, so none of this has to be taken on trust.