AI News

The most expensive item of Double 11 is a 24.5-yuan Token package

TMTPost · AGIUpdated Nov 16 18:42

AI summary

For the first time, Tmall Double 11 has put large-model Tokens on the shelves as monthly, quarterly, and annual cards. Seven vendors including MiniMax, Zhipu, and Alibaba Cloud Bailian have joined, and the MiniMax Go package costs as little as 24.5 yuan after a 50% discount. However, each vendor's metering is split into three non-convertible types—Credits, points, and progress bars—and the same package's actual consumption varies greatly across different tasks, making price comparison untenable. In the same period, the platform canceled cross-store full-reduction discounts and switched to official direct discounts.

Body · AI translation

By Tingfeng Yima

At 8 p.m. on October 20, Tmall's Double 11 officially opened sales across all categories.

Among the 72 million products, there is one thing that cannot be seen or touched, and has no tracking number.

It is called Token, with the Chinese name ciyuan (word element), and it is the unit of measurement used when AI talks, solves problems, and draws images.

This year, for the first time, it was placed on the Double 11 shelves, sold as monthly, quarterly, and annual cards, with the cheapest tier at 24.5 yuan.

For 17 years, Double 11 has taught Chinese people, step by step, how to calculate full-reduction discounts, meet spending thresholds, and compare final prices.

This year the platform canceled cross-store full-reduction discounts, yet the first new product it put on the shelves is a commodity whose price the entire industry cannot calculate clearly.

(1) An invisible product goes on the shelves

A media outlet once recorded a shopping cart like this: a foldable-screen phone, a year-long cleaning service package, and a large-model Token package.

The cart's owner was a programmer in Hangzhou, who said his biggest trophy this Double 11 was a cleaning lady who can come to his home all year round.

On Double 11 eighteen years ago, the shopping cart contained only the first kind of thing.

The word ciyuan was officially named by the National Data Administration only in 2026.

It is neither a word count nor a number of times, but the smallest unit that a model splits information into when processing it.

When you send a passage in, the model first cuts it apart, numbers it, calculates probabilities, and then spits it out piece by piece.

Every piece cut out is billed once.

So it has no weight and no volume, yet it can be ordered and invoiced, and it is just a number.

This year's numbers are 1 million global brands and 72 million products, both setting new records, with in-stock sales running from 8 p.m. on October 20 all the way to 23:59 on November 13.

Among these 72 million products, there is a category that has no color, no size, does not enter warehouses, and is not shipped; after ordering, the quota is directly credited to the account and is used up within the package's validity period.

Tmall spent two months preparing to put Token on the shelves.

Starting in September, Zhipu, Alibaba Cloud, and Qianwen Office successively opened official Tmall flagship stores.

On October 13, MiniMax opened the first brand official flagship store, with three subscription tiers on the shelves at official website prices of 49 yuan, 119 yuan, and 469 yuan, all at half price during Double 11, with the lowest tier at 24.5 yuan.

By the October 15 press conference, there were already 7 model vendors onboard: Alibaba Cloud Bailian, MiniMax, Qianwen Office, Moonshot AI, Qoder, Zhipu, and StepFun.

According to a Time Finance report, Tmall President Jialuo's exact words were that almost all large-model companies are selling Token on Tmall, and before official sales he expected two more leading companies to join.

His goal was more specific: to assemble the Token sales entrances of the full matrix of Chinese large-model companies.

Another set of figures was given by Cheng Yuan, general manager of 3C Digital: this Double 11 concentrated the launch of more than 1,000 new products, of which more than 70 are AI personal computers preinstalled with Qianwen Office, involving brands such as Lenovo Zhaoyang and Lecoo, ASUS, MSI, and ThundeRobot, and the platform lists them as a separate new category with the code QAIPC.

To translate: one is selling electricity, and the other two are selling appliances with motors installed.

The platform has a clear idea of who the buyers are.

Jialuo divided the target groups into four categories: people who are studying, family users, people doing research, and one-person companies or three-person small teams.

These people were already steadily using models, but were scattered across various platforms paying by consumption, with no single place where several could be compared side by side.

The change is also very intuitive in the buying experience.

According to a Beijing Daily reporter's test, when searching for any model brand on Tmall, the page that pops up at the top of the search results is almost identical to a mobile phone credit recharge: choose a package, enter a phone number, pay, and the quota arrives in seconds.

In the product review section, some wrote "arrives very fast, easy to get started," while others wrote "the usage instructions are very detailed."

The number of units sold in several flagship stores ranges from dozens to several thousand.

There is another reason beyond technology why Token can be put on the shelves.

It has no logistics, no inventory, and no installation or after-sales service, and its marginal delivery cost is nearly zero.

For the platform, this is a new supply that requires almost no fulfillment cost.

Hardware is keeping pace: Cheng Yuan mentioned that these more than 70 AI personal computers were jointly defined by chip makers, model providers, and the platform, with Intel and AMD on the chip side.

Selling quotas, selling devices, and selling preinstalled capabilities—three lines landing together in the same major promotion.

To buy, one had to find a technology vendor, sign a commercial contract, configure a key in one's own backend, and then settle by API call volume, paying for what was used.

After being moved onto e-commerce shelves, the form changed accordingly.

No longer a string of keys in the backend, but cards bought by the month, quarter, or year, as well as bundle packages combining several services.

Jialuo found a reference for this kind of bundle: beauty sets.

His phrasing was to let stocking up on Token be like girls buying beauty sets on Double 11, cheaper than buying individual items.

So on the shelves appeared study packages, family packages, research packages, and one-person studio packages, grouped by audience and equipped with the few items each would find useful.

Zhipu's tiers are divided into a personal version and a team version standard seat, compatible with more than 20 mainstream Agents such as ZCode, Claude Code, and Codex.

MiniMax's M Plan is divided into three tiers, Go, Explore, and Build, with quotas progressing at roughly 1x, 3x, and 12x.

Packaging expensive compute into standard SKUs matters because it can now be traded, repurchased, and compared side by side.

On this path, telecom operators moved earlier than model vendors.

Starting in May this year, the three major operators successively launched token packages, and a single package can usually call multiple domestic models.

China Telecom's Lite version costs 9.9 yuan for 10 million Tokens.

According to a Daily Economic News report, these standardized compute refuel packs cut the cost of generating one minute of AI video from fifty or sixty yuan to about 1 yuan.

The significance of this path is that tokens have been placed into one of the consumption habits users are most familiar with.

It is not about first learning how to call an API, but about buying compute the way you top up a phone bill.

The speed at which the barrier has come down can be seen clearly from a larger set of numbers.

Data from the National Data Administration shows that China's average daily token calls rose from about 100 billion in early 2024 to 100 trillion by the end of 2025, and further surpassed 140 trillion in March 2026, a more than 1,000-fold increase in two years; the scale in March 2026 was about 1,400 times that of early 2024.

The National Data Administration also confirmed the Chinese translation of Token during the same period: ciyuan.

In its research report on the token economy, the China Academy of Information and Communications Technology cited an IDC forecast: in 2026, China's MaaS market will consume 40 quadrillion tokens for the full year, compared with 1.944 quadrillion in 2025, a more than 20-fold increase in one year.

The same report also noted that token consumption is moving from human-computer interaction toward large-scale autonomous activity, and that the tokens consumed by an agent in a single task can reach 10 to 100 times that of an ordinary ChatBot.

The change in magnitude is also rewriting how this business is priced.

In its report, the China Academy of Information and Communications Technology mentioned that the focus of industry development has shifted from model training to inference services, and that large models have officially entered the inference era.

Nvidia redefined the data center as a token factory and upgraded cost per token and token throughput per watt into core industry competitiveness metrics.

In translation: the compute business has begun to be priced by output, no longer by equipment.

The China Intelligent Internet Development Report (2026) released by People's Daily Online contains another set of figures: according to data published by Cloudflare Radar in July 2026, machine traffic already accounted for 59.2% of China's HTTP requests, surpassing human traffic.

The same report shows that in 2025, projects of the large-model application type accounted for 58%, becoming the industry mainstream.

Machines have begun to access the internet more frequently than people, and every visit they make burns tokens.

From 100 billion to 140 trillion, tokens completed an identity switch in just over two years: from a cost item for enterprises to a commodity consumers can stock up on.

A Beijing Daily reporter bought from several flagship stores one by one and found that on the same shelf, there were at least three measurement standards, and they were not equivalent to one another.

Alibaba Cloud's limited-time 39-yuan Lite package has a monthly limit of 11,500 Credits.

The MiniMax Go package, originally 49 yuan and 24.5 yuan after discount, does not give specific token usage parameters on the page; users can only go back to the console to look at a usage progress bar.

The 118-yuan Zhipu Lite monthly card has a quota of 10,000 points per week.

Credits, progress bars, points: three companies, three ways of saying it, and none of them tells you how to convert between them.

What is even harder is the time window.

One vendor's non-video models are controlled by a 5-hour window, with a limit of 2,000 points every 5 hours.

In other words, you are not buying a month's total, but the sum of quotas across several windows within a month; exceed the limit and you have to wait, or pay more.

The complexity of conversion directly destroys price comparison.

A case of bottled water can be compared across three platforms because what you buy in all three places is the same case of water.

But which model's capabilities a Token package contains, which Agents it is suited for, and in which scenarios it is cost-effective—no unified standard can line it up in a row with the one next door.

Price comparison presupposes homogeneity.

The same phone sold at different prices on three platforms, with consumers choosing the lowest, works because the goods are the same.

Model services have no homogeneity.

The same sentence, "help me write a weekly report," can produce very different output quality, response speed, and whether it can be directly plugged into an Agent, depending on the model.

The tokens consumed by one agent task may be 10 to 100 times that of an ordinary conversation.

This means that for the same 24.5 yuan, a user who only chats can use it for a month, while someone running code with an Agent attached may not last a week.

There is another contrast that can help calibrate your sense of units.

9.9 yuan buys 10 million tokens, which sounds like a scary number.

But ask the reverse question: how many does it take to generate a research report dozens of pages long, run a full data analysis, or have an Agent revise an entire project?

No one answers this question on the product page.

Inconsistent units are not just an oversight.

Each model has a different capability structure, and the cost curves for long text, multimodality, and Agent calls also differ; billing by token, by points, and by number of uses rests on three different cost models.

A unified standard would mean laying out your own cost structure for competitors to see, and no one is willing to move first.

The blank space in price comparison will not be filled by anyone in the short term.

There is also one detail that is easy to overlook.

When the quota runs out, users usually do not face a shutdown; instead, they face throttling, queuing, or buying a temporary top-up pack.

The unit price of these top-up packs is generally higher than the average price within the plan.

If you insist on working out the math, there is only one way: run it yourself.

Pick a task you do most often, such as asking AI to write a 3,000-word report, or asking an Agent to finish revising a front-end page. Note the quota once before doing it, and once after.

The difference between the two numbers is your true per-task cost.

Multiply it by your monthly task volume, then compare that against those plan price tables; only then does the math hold up.

This is like going to a gas station to stock up on fuel: the station tells you the price per liter, but not how many liters your car burns per 100 kilometers.

What is cheap is not the Token; it is the bill you cannot calculate.

(4) Full-reduction exits as platforms settle an old account

In the same major promotion in which Token was listed, the platform cut the playbook it had used for more than a decade.

Taobao and Tmall officially canceled the cross-store full-reduction and order-padding mechanisms, and changed the core playbook across all stages to official direct reductions.

According to Jiemian News, official direct reductions take effect at the store level. After merchants sign up, products in participating categories in the store are automatically included. The minimum discount rate is divided into three tiers by category—1%, 5%, and 15%—to accommodate the profit margins of different categories.

On top of this floor price, store coupons, category coupons, the 88VIP 5% discount, and super red packets can be stacked, creating a ladder of direct reductions as the base and benefits layered on top.

JD.com retains cross-store full reductions, mainly promoting 50 yuan off for every 300 yuan spent, while also pushing official direct price cuts; Douyin offers an official 15% direct reduction plus a direct price cut on single items.

The direction is the same; the intensity differs.

The order-padding that was cut is essentially a form of price discrimination: people willing to spend half an hour doing the math get a low price, while those too lazy to calculate pay the original price.

With two sets of prices on the same page, consumers choose sides themselves.

The platform abandoned this mechanism not because it was complicated, but because its side effects had outweighed the benefits.

Order padding created a large number of unplanned orders: consumers casually added items to cross the threshold, then applied for a refund on the filler item after paying.

Public data shows that during Double 11 in 2025, the comprehensive e-commerce return rate was 61.5%, with the return rate in apparel livestreaming exceeding 80%.

The return rate for women's clothing livestreams has long hovered around 80%, with some merchants reaching 90%; two years ago, the average return rate for this category was still around 30%.

The platform's reported GMV rose, merchants' stocking costs rose, the volume of logistics parcels rose, and real demand did not move at all.

Once growth slows, the fulfillment cost of every order becomes glaring.

A report released by Syntun shows that during Double 11 in 2025, total online e-commerce sales across all platforms were 1,695 billion yuan, up 14.2% year on year, while in 2024 the growth rate of this indicator was 26.6%.

Of that, comprehensive e-commerce sales were 1,619.1 billion yuan, up 12.3%, and instant retail grew 138.4%.

Data from the National Bureau of Statistics shows that from January to August 2026, national online retail sales of goods grew 4.3% year on year, with food up 15.9%, clothing up 4.9%, and use goods up 1.1%.

Clothing and use goods are barely growing at all.

The battle lines are also lengthening in parallel.

In the previous major promotion, Douyin ran for 57 days, JD.com for 37 days, and Tmall for 31 days.

The longer the cycle, the flatter the single-day peak, and the weaker the concept of Double 11 day itself becomes.

Regulation also tightened at the same time.

According to The Paper, on October 10 the Ministry of Commerce and six other departments jointly issued a notice on implementing the five-excellence action for quality e-commerce, explicitly correcting disorderly competition such as automatic price matching and lowest prices across the entire network, and regulating platform commission cuts and merchant rating rules.

When the price war is put in a straitjacket, platforms have only two cards left to play: make the rules simpler and make the shelves wider.

Hence the two sides of this year's event: cutting full reductions on one hand, and listing Token on the other.

The former is settling old accounts; the latter is finding new ones.

Membership is another pillar of this new stratification.

Tmall previously disclosed that more than half of the business of leading brands is driven by 88VIP.

Data from the Taobao Live New Product Festival shows that among purchasers of new products with monthly livestream sales exceeding one million yuan, 88VIP members accounted for more than 60%, and their average order value was 21.6% higher than that of overall new-product buyers.

A survey shows that 79.4% of consumers ask AI before shopping, and AI's influence on consumer decisions has reached 43.7%, surpassing traditional media advertising for the first time.

Rules subtract, shelves add; having finished selling things that can be calculated, they begin selling things that cannot.

(5) Before stocking up, answer three questions

Back to that 24.5-yuan product.

First, to be clear about the prices themselves: 24.5 yuan is MiniMax's first-order 50% discount price, 118 yuan is Zhipu's monthly fee, and 39 yuan is Alibaba Cloud's limited-time price.

The three differ entirely in original price, cycle, and quota basis; they appear on the same page only because they are all called plans.

To judge whether it is worth stocking up on, the discount intensity provides almost no information; you have to answer three questions yourself.

Question one: Can you use it?

For people with clear scenarios, stocking up on Token means locking in the price of spending that has already occurred.

For people who merely think they might use it later, what they buy is an imagination, and imagination usually does not materialize before the plan expires.

Every provider has a free tier and trial quota; use up the free tier first, then discuss whether to stock up.

Question two: Can you calculate it?

Does this provider tell you how many tokens a task equals, how much money that equals, and how tokens, points, and number of uses convert into one another.

If the conversion is transparent, you can compare prices; if the conversion is a black box, first buy a small monthly card for one month, run it for a full month to see the real consumption curve, and then decide whether to go for an annual card.

Across from 24.5 yuan is 49 yuan; across from 118 yuan is how much it will still sell for next year.

The amount locked in by an annual card—can you really use it all up in 12 months? The unused portion—will it be zeroed out on the expiration date, or extended?

There is also a category of traps written into the terms, not on the price tag.

Some plans have quotas that reset monthly and are forfeited if unused; top-up packs often have their own separate validity periods; and after an annual card is sold, most do not support mid-term refunds.

Spending one minute before ordering to clearly check the validity period and refund rules is more practical than stacking another coupon.

Asking these points clearly before ordering is worth more than any coupon.

The answers to three questions determine how to buy: if all three are yes, you can confidently stock up on an annual card and lock in the lowest price.

If two are yes, first stock up on a monthly card to test the waters, and run it for a month to see real consumption.

If only one is yes or none are, first milk the free tier dry, and do not pay for something you might use in the future.

If you decide to buy, the stacking path is like this.

The first wave of 88VIP shopping coupons was issued on October 13, 7 coupons in total, starting at 12% off; they can be redeemed after sales officially start at 8 p.m. on October 20, and they support choosing the denomination to inflate: spend 480 minus 60 can be inflated into any of three tiers—spend 600 minus 75, spend 800 minus 100, or spend 1,000 minus 125.

After brand discounts are stacked with platform shopping coupons, the overall strength of this Double 11 is as low as 28% off, a peak in recent years.

Buying an AI computer preloaded with a large model lets you stack another layer of national subsidy: home appliances and digital products are subsidized at 15% of the actual paid price, with per-item caps of 1,500 yuan and 500 yuan respectively.

Remember the time window: sales officially start at 8 p.m. on October 20, and in-stock sales run until 23:59 on November 13; both 88VIP coupons and platform coupons are redeemed within this window.

There is also one plainest piece of experience that applies: the best deal of Double 11 is always what you were going to buy anyway.

Tokens stocked up just for the sake of stocking up are waste no matter how cheap.

Final words

For 17 years, Double 11 has been doing the same thing: calculating the price of a product to two decimal places.

Cross-store full reductions, deposit inflation, final price—all the gimmicks point to the same goal: making you certain you got a bargain.

This year, the platform personally dismantled this algorithm.

In the same major promotion that dismantled it, the shelves gained a product that cannot be compared by price.

Starting at 24.5 yuan, 7 manufacturers on the same stage, and a market with 140 trillion daily average calls endorsing it.

When even whether a product is worth it cannot be calculated, discounts lose their meaning, because a discount must be relative to a known value to hold.

The most expensive product of Double 11 has never been the one with the highest listed price.

It is the one where, when you press the pay button, you still do not know how much of it you bought.

Source

TMTPost · AGI