450% is the month-on-month jump in Tmall's token category transaction value on the first day of the Singles' Day promotion, and not one of those prices is comparable across bundles. As of November 11, 2026, AI tokens have become a product category on Chinese marketplaces, yet they are still merchandised like phone credit: packs whose usefulness depends on the model behind them. JD.com reported token package sales up 167% year over year in its opening phase (Tech360, citing SCMP). On November 2, nearly 100,000 QwenWork monthly memberships sold in 30 minutes at 0.01 yuan, representing about 2 trillion tokens (CGTN).
A new category with no unit of measure
Tokens are the units a model uses to read and write text. The same Chinese sentence can be split into different token counts by different models, so identical quotas do not represent equal compute. That breaks the first thing a marketplace needs, which is a comparable price per unit of the thing actually being sold.
The discount ladder compounds the problem. A monthly subscription normally listed at 28 yuan was sold for one fen, which is an acquisition cost rather than a price. Subscriptions sold, tokens contracted and capacity actually used are three different numbers, and only the last one predicts revenue.
Three numbers that look like demand
Platform disclosures tend to lead with the first two, because both are large and both are verifiable. Neither says anything about renewal at full price, which is the only test of whether a token package behaves like a product rather than a promotion.
| Metric | What it shows | What it does not show |
|---|---|---|
| Subscriptions sold | Acquisition volume | Renewal at full price |
| Tokens contracted | Capacity attached to the sale | Capacity actually used |
| Category GMV growth | Promotional pull | Unit economics |
JD.com pushed token packages from more than ten model providers into a November 2 livestream, and sales of large-model token packages rose 20-fold within two hours on one channel (CGTN). That is a distribution result, not a pricing result. It tells a brand that livestreaming can move AI subscriptions; it says nothing about what buyers will pay once the discount ends.
What a buyer can actually compare
Normalise before you compare
The practical fix is to convert every package into a price per thousand characters of the same reference text, generated by the model in question. That number is comparable across providers; a price per token is not. It also exposes packages that look cheap because their tokenisation is coarser.
Watch the expiry window
Longer expiry windows encourage stockpiling and flatter the sales number in the month of the promotion. A buyer comparing two packages should compare price per character over the period they will realistically use them, not over the maximum validity printed on the package.
The governance question
Price transparency in this category is not only a platform problem. If a marketplace lists token quotas as the headline and the underlying unit is not standardised, the comparison shoppers make is systematically wrong. The published numbers describe a promotion, not a normal purchase, and the gap between the two is the whole story.
A brand selling AI services should expect the same scrutiny that physical goods get: a stated unit, a stated validity, a stated refund path. Without those three, the category stays promotional, the price gap stays unmeasurable, and every subsequent discount has to be deeper than the last.
FAQ
Why can't token prices be compared directly?
Because models tokenise text differently. The same sentence can produce different token counts, so identical quotas do not represent equal compute.
What should a buyer compare instead?
Price per thousand characters of the same reference text, plus the validity window and the refund path.
Is the discounting sustainable?
Only if renewal at full price follows. Subscriptions sold and capacity used are different numbers, and only usage predicts revenue.
