Turn just-enough intelligence into a viable business.

On October 6, OpenAI announced a set of mathematical research completed by its internal cutting-edge models. It also disclosed a quite interesting figure: to obtain one research result, the average computational consumption is equivalent to about three hours of thinking by ChatGPT Pro.

The next day, Anthropic released its new-generation small model Haiku 5.5, stating that it is suitable for a large number of repetitive tasks such as summarization, classification, and database query; also on the 7th, OpenAI rolled out GPT-6 to more ChatGPT users, while free users are provided with the lighter Luna. Cutting-edge models continue to explore scientific challenges, while daily requests are increasingly handled by lower-cost capabilities.

This same divergence path has also appeared in the products of major Chinese tech companies. Alibaba talks about the future of superintelligence, while Qianwen Office is busy answering operation questions from store employees; Tencent invests in models and computing power, and also lets WorkBuddy undertake work tasks, and enables Marvis to clean up computers and find files for individual users; ByteDance integrates Doubao Work into Feishu, allowing it to access authorized team materials and return editable documents.

Loading sufficient intelligence into familiar software is not as exciting as sprinting for capability records, but it determines how these companies digest the increasingly heavy AI investment.

This type of intelligence is not a height-measuring competition to see who can solve math problems first, but an art of sufficiency, which depends on how far manufacturers can deliver: to well meet user needs, and achieve surplus under the pressure of heavy asset investment, or even better, achieve surplus with sustained growth.

After all, what most people encounter every day are not mathematical conjectures: they need to know how to post a poster, where the contract is, why the computer is lagging, and what was agreed in last week's meeting. So, superintelligence may represent the imagination of this era, but AI ultimately needs to be implemented as a viable business in the process of technology diffusion.

01

Alibaba CEO Wu Yongming called AGI a starting point at the Yunqi Conference in September, talking about superintelligence capable of self-iteration. Strictly speaking, superintelligence refers to intelligence that far exceeds the best human performance in almost all important fields, which is still a future vision; what enterprises can purchase and compare prices for today are cutting-edge models, and the two cannot be equated.

The practical problem faced by product managers is: how many requests a year actually require the strongest model available today? The time consumption of OpenAI's publicly released mathematical research this time cannot be directly converted into the dollar cost of each question, let alone be used to price superintelligence. It at least shows that in cutting-edge exploration, multiple attempts, verification and long hours of thinking are inherently part of the work. Enterprises will not pay for model queries to find a piece of information in the same way they pay for mathematical research.

According to the public API pricing, the standard tier of GPT-6 Astra costs $50 per million output tokens, and GPT-6 Luna costs $0.5, a 100-fold difference on the surface. For Haiku 5.5 released on October 7, for prompts no longer than 100,000 tokens, the pricing is $0.1 and $0.5 per million input and output tokens respectively; Anthropic stated that the new model is suitable for high-frequency, cost-sensitive tasks.

Different models have different token consumption, output quality and rework times for the same task, so it is impossible to judge which one is more cost-effective just by looking at the price list. But when an ordinary query is scaled up to tens of millions of calls per day, the price gap will be reflected in the budget.

Assuming that 10,000 employees get 1,000 tokens of answers per day, the total output alone will reach 10 million tokens a day. At the public standard price of Astra, that is $500, while at the price of Luna, it is only $5. This does not include input, document retrieval, tool calls and retries after failures, nor is it the actual quotation for any office product purchased by enterprises.

Of course, in real scenarios, the token consumption of a 10,000-employee company cannot be only this amount. This means that using AI is not like buying an iQiyi membership where you can top up as you like. In fact, even individual users will carefully judge their actual needs and spend according to their income when purchasing AI product memberships.

Fortunately, the models themselves are getting cheaper. Anthropic said that Haiku 5.5 is about 75% cheaper on average than the previous generation when calculated by actual tasks, and at the same time halved the price of cache reading for Sonnet 5.5; OpenAI also previously lowered the prices of Luna and Terra. Tasks that could only be handled smoothly by expensive models last year can now be completed by smaller models.

Therefore, sufficient intelligence does not refer to a single cheap model, but the ability to allocate computing power according to tasks: locate documents accurately first, assign simple tasks to fast models, upgrade to higher tiers for difficult tasks, and reuse repeatedly read materials as much as possible. If the cheap model leads to repeated retries and a large amount of manual rework, the apparent token discount will be offset.

In September, the Tencent Marvis team stated in an exchange that after indexing local files and videos, improving cache hit rates, and selecting models by task difficulty, the cost per daily active user dropped to about one-tenth of that in the first month of launch in May. Not long ago, Qianwen Office also announced that dedicated models reduced the average token consumption by 75% in office tests.

These two sets of data point to the same process that determines whether a product can be deployed on a large scale: make one small task cost the price of one small task.

Price reduction also has its downside. As model suppliers sell the same capability at lower and lower prices, application developers naturally have lower costs, but also lose some reasons to charge solely for model capabilities. Customers pay for automatically sorting a meeting record today, and tomorrow collaborative software may treat it as a free feature. Only when the product can find the correct company information, complete the approval process, and find the responsible person when an error occurs, can the capability stay on the procurement list. The more popular scalable intelligence becomes, the more solid the supporting services outside the product need to be.

Alibaba disclosed capital expenditure of 67.678 billion yuan in the last quarter, and Tencent's capital expenditure was 52.8 billion yuan, up 75% and 176% year-on-year respectively. The calibers are different, and there are long-term investments that do not belong to office agents.

They still remind people that the expenditure on computing power, models and products needs to be supported by more businesses. Turning intelligence suitable for daily tasks into commonly used products is a way to share the huge investment.

But whether this path leads to incremental business still depends on how contracts are implemented in actual execution, and whether the unit economics (UE) can be maintained at a healthy level.

02

On October 7, OpenAI added an "intelligent interface" to ChatGPT: when users raise a question, the system can directly generate buttons, forms, charts, and even a temporary usable small tool. Users may no longer need to search for a calculator in the app store first, or learn the menu of a certain software before starting work.

It is not proof that all independent applications will disappear, but it puts the pressure on traditional applications in front of everyone: if users first state their purpose and let AI decide which tool to call, the software that originally occupied the home page and icons may retreat to the background.

For Tencent's Appbao, its past value lies in delivering applications to users, and the platform realizes its commercial value as a distribution channel. It can be said that the more applications users need, the more valuable the channel is. But what if users' demands are converged?

This is the background why the same team developed Marvis. Marvis starts to undertake users' specific task demands. For example, when a user asks "Why is my computer so slow", Marvis helps the user optimize the computer system; when a user asks "Help me find that video", Marvis can accurately locate the relevant material in the computer's documents.

Developers will also face the same problem. In the past, developers turned a function into an App, and waited for users to download it after it was launched; now the function may be first discovered by AI, and then presented in the form of a call, a skill or a transaction. Marvis hopes to integrate professional software capabilities. If this path is successful, the "distribution" of Appbao will become service matching in tasks.

If only free system operations are frequently called, the platform that occupies the entry may not necessarily obtain enough new revenue.

Similarly, office applications such as DingTalk and Feishu are also facing the challenges brought by AI. A large number of task-oriented applications and agents have emerged in the market.

However, these third-party agents can write a beautiful weekly report, but they do not have the employee identities, approval processes, group chats, calendars and a large number of non-public documents that are precipitated in DingTalk and Feishu. Even if they get authorization, they still need to re-solve the problems of data access, permission inheritance, update and delivery. They don't even know which provinces are included in the "East China Region" mentioned by the company, or who approves the contract.

Qianwen Office launched "Enterprise Context" in September, allowing enterprise messages, knowledge and business data to be accessed as needed for tasks; it can access DingTalk, as well as Feishu and WeCom. Feishu, on the other hand, opens up organizational permissions, documents and collaboration environments for Doubao Work, so that the generated content can be returned to the team for further editing. The two companies take different approaches, but both are striving to become the place where employees assign tasks.

Take the milk tea chain brand Guming as an example. It launches new products very quickly, and the rules about how often to replace filter elements, how to make tea soup, and how to deal with machine failures may be scattered in training materials and group chats. When a store employee asks a specific question, the general model does not know the specifications of the day, and Qianwen Office organizes these materials to find answers according to permissions.

This practical path shows that whether an intelligent service can be "sufficient" depends not only on the model, but also on whether the enterprise has provided the correct information to it.

This also makes the old assets of large tech companies valuable again. Models can be purchased by others, but organizational relationships, permission systems, business processes and customer service experience cannot be replicated overnight. However, accessing these things itself costs money, and maintaining incorrect permissions, expired materials and operation records also costs money.

In the past, software sold a wrench, but today it promises to tighten the screw for users; the responsibility is far more complex than the interface.

03

Selling a wrench is one price, while tightening a screw brings further value, and this logic is very reasonable. But the problem is, are enterprise users and individual users willing to pay more budget for this?

After all, sufficient intelligence is still intelligence, and there is still a large amount of resource investment behind it. The AI upgrade of old products and old teams is undoubtedly a battle to retain users, but if it cannot leverage more budget, it cannot be called a good business.

The charging problem is different for consumers and enterprises. Marvis first needs to answer: are people who are willing to let AI manage their computers also willing to pay for it? Qianwen Office and Doubao Work need to answer: enterprises have already paid for DingTalk and Feishu, why do they need to add an extra AI budget?

The relevant person in charge of Marvis previously stated that charging will not start until the product value is sufficiently mature; some users have complained that they cannot make a purchase after the free quota is exhausted. The fact that some people want to buy more usage is a lead, not a stable payment rate.

Tasks such as cleaning disks, finding files and installing software are usually free in past systems and tools. If Tencent turns these basic actions into expensive subscriptions, users have alternatives; if they are permanently free, heavy usage will continuously consume inference costs. The reduced unit cost has bought it trial time, but does not automatically generate revenue.

Tencent is testing several charging positions: let professional developers integrate their capabilities into the manager, and users pay for the skills that are really useful; strive for pre-installation on PCs, and cooperate with NAS and Mini PC manufacturers; encapsulate processes such as inspection for industry customers.

But in general, for the PC manager product, charging individual users does not have optimistic prospects, because this business has a very long history of free use.

The enterprise side at least has mature procurement practices. The enterprise version of Qianwen Office is sold by seats, with 2000 points per seat per month, shared by the organization, and additional usage can be purchased through point packages. Letting companies buy a fixed number of seats first, and then supplement the usage when they use more, builds a bridge between the old software contracts and the model bills that fluctuate according to call volume, but the challenge is how to convince enterprises to widen this bridge.

The shared quota seems to be a billing detail, but it actually adapts to the extremely uneven usage in the office. The finance department performs batch verification at the end of the month, the legal department reviews contracts intensively before transactions, and the sales department processes a large number of inquiries during peak seasons. It is impossible to require each employee to just use up the fixed tokens every day. The unified point pool allows enterprises to allocate quotas to busy employees, and the platform can continue to charge when the usage exceeds the limit. However, once tasks are truly handed over to agents, the cost depends not only on the number of employees, but also on how many files it reads, how many rounds of processes it runs, and which tier of model it calls.

Purchasers will require a predictable upper limit and verifiable usage records, while suppliers need to prevent a fixed annual fee contract from being exhausted by a small number of heavy tasks.

A company originally spends money on collaborative software every year. Now the supplier adds AI to it. If customers only move the same budget from the old version to the new version, the platform increases the cost of model inference, data access, permission management, audit and customer delivery, but the revenue may not necessarily increase.

To obtain incremental budget, AI has to enter the work that was difficult to charge for in the original office software. For example, the warehouse logistics packaging company Fulika lets digital employees handle inquiries, quotations and approvals; in legal services, AI office applications are used to verify materials in batches. Only when business departments are willing to allocate funds from sales, operation or professional service budgets, AI office will not be just a budget reallocation within the IT department.

What Qianwen Office is fighting for is exactly this position: let the organization hand over a process that was previously unaffordable or too slow to it. The success or failure of this business will not be written on the parameter list, but in the contract of the next year: how much new budget is obtained, how many things are done for customers, and how much cost is paid.

This article is from the WeChat official account "FunTalk" (ID: iFuntalker), written by Lin Shu, and authorized for release by 36Kr.