Tmall Selling Tokens: Like Prepaid Phone Bills or Blind Boxes?

I spent two days testing the Zhipu AI Tmall flagship store. Conclusion first: I don't recommend beginners jumping in directly; it depends on the situation. If you already know which models you're running and roughly how much quota you need per month, using it as a payment gateway is quite convenient. But if you've just heard that AI can chat and want to casually top up, it might feel like opening a blind box.

An engineer told the media that buying Tokens will soon be as smooth as topping up your phone bill.

That sounds nice. The first instinct of an investigative journalist is usually: Is this data source reliable? Is the purchase smooth, or is it the subsequent consumption and after-sales service?

Let's explain Tokens first. In Chinese, they are often called "ci yuan" (word tokens). Large language models chop text into small chunks for billing. You input a segment, it outputs a segment, and both deduct from your quota. Buying Tokens isn't like buying a monthly video streaming subscription where you watch unlimited content; it's more like buying a prepaid card—whatever is on the card gets used up, then you recharge. Different models, different lengths, and different tool calls may have different deduction rules.

On September 2nd, Zhipu AI opened a store on Tmall. The only items listed were four Token packages at different price points. When I opened the flagship store, the homepage was very clean, with four package cards side by side. There were no sales figures, reviews, or user photos typical of physical goods stores, indicating low activity—it felt like moving a developer backend onto an e-commerce shelf. I picked the cheapest tier, added it to the cart, and paid. This step was indeed hassle-free, similar to buying a membership.

The bottleneck appeared after payment. I wanted to confirm if the quota had arrived, if I could specify which model to use it for, and if refunds were possible. The product page wasn't comprehensive, so I contacted customer service. They said usage could be checked via official portals, and specific consumption follows the model's billing rules. Is this data source reliable? I still had to go back to my usual terminals, browsers, and model dashboards to verify things from multiple angles.

I also tried an interview summary task. I asked the model to compress a long transcript into an outline, then revised it twice, going back and forth several rounds. Subjectively, the quota decreased. However, in my tests, it didn't provide itemized deductions like a phone bill; there was only a change in the total balance. For ordinary users, this isn't reassuring enough. It's hard to judge which step cost more—was the model expensive, did it process too much content, or were there many tool calls?

There were pleasant surprises too. Tmall's payment and after-sales systems have a low psychological barrier for users, at least friendlier than binding cards and configuring API keys (authorization codes) in a developer backend. For many non-technical readers, being able to complete a purchase on a familiar e-commerce page is itself a form of popularization.

But I still don't recommend novices stocking up on large packages right away. Token consumption isn't fixed like a phone plan. When the page doesn't clearly explain the consumption curve, users easily mistake "topping up" for "buying." For deep investigators, this ambiguity is the most troublesome because accountability later gets scattered between the store, the platform, and the model service.

E-commerce shelves naturally emphasize price, not boundaries. The four packages solve "how to buy," but haven't fully solved "how to use, how to stop, and how to refund." When I wrote about Agent foundations before, I said the core of competition is an auditable and accountable operational system. Looking at retail computing power now, it's the same. After verifying from multiple sources, I feel it's more like a convenient checkout counter for users who already understand the tech, not an educational entry point for beginners.

Action Advice: Don't stock up on large amounts yet. If you really want to try it, buy the lowest tier, run a fixed task, record the number of inputs/outputs and remaining quota, and confirm refund policies, invoices, and validity periods before considering renewal. It suits people with clear usage needs, not beginners who just want to taste-test.