DeepSeek Brings Spring to AI Application Entrepreneurs

Original Title: "DeepSeek Brings Spring to AI Application Entrepreneurs"
Original Author: Dongcha Beating

On July 31, the official version of DeepSeek V4-Flash API was launched for public testing.

The cost is 1 yuan per million input tokens and 2 yuan per million output tokens. After cache hits, the cost drops to only 0.02 yuan per million input tokens. It supports 1 million context, reasoning modes, tool calls, and Responses API, with a maximum concurrent connection limit of 2500 per account.

DeepSeek will also adopt peak and valley pricing, doubling prices during peak hours. Even so, the prices remain surprisingly low.

According to rough calculations provided by DeepSeek, one Chinese character consumes about 0.6 tokens. One million tokens can read approximately 1.67 million Chinese characters. Looking solely at input costs, reading about 1.67 million Chinese characters costs only one yuan.

In the past year, the large model industry has become accustomed to longer contexts, more complex reasoning, and increasingly capable agents. With each model update, discussions often revolve around who has surpassed whom, how much benchmarks have improved, and how far we are from AGI.

For AI application entrepreneurs, there is an even more pressing question:

How much does a task actually cost?

Model capabilities determine what products can theoretically do, while calling prices determine how much capability teams are willing to integrate into their products.

This afternoon, I had a lengthy discussion with ChatGPT about this. I said that the current business model for AI applications is essentially reselling tokens.

It tried several times to help me rephrase it more elegantly, suggesting terms like "retail of reasoning capabilities" or "intelligent service packaging." I told it not to sugarcoat this business; reselling is reselling, and just because it sounds harsh doesn't mean it's wrong.

AI applications seem to be the latest entrepreneurial direction in recent years, but fundamentally, they are based on an old business model:

Purchasing, processing, pricing, and then selling the goods.

Many industrial changes often begin with a sudden drop in purchase prices.

AI Applications Are a Business of Reselling Tokens

In the past, people were accustomed to understanding AI applications through the framework of internet products or SaaS.

Once software development is completed, it can be distributed to a thousand people or a million. Adding a new user will certainly incur bandwidth, storage, and service costs, but these costs usually get gradually diluted as the scale expands.

AI applications are different. Every time a user generates an image, writes a report, completes a piece of code, or has an agent execute an additional step, the application company must call the model again. Every time a user presses a button, upstream costs are incurred again.

Thus, the basic business model for most pure software AI applications is quite simple:

Purchase tokens from model companies, package general model capabilities into writing, programming, design, customer service, video, or agent products, and then sell them to users through memberships, points, and pay-per-use fees.

For example, Poe directly stated its pricing goal when launching its API: the pricing of additional points is based on the prices charged by the underlying model suppliers.

Upstream charges for models and tokens, Poe converts them into points within the platform, and sells them to users along with a unified entry point, payment, model switching, and development interfaces.

Some companies claim they do not sell tokens but sell results. The pricing for the AI customer service Fin designed by customer service software company Intercom is $0.99 per completed result. A result can be counted when the user confirms that the issue has been resolved, there are no further requests for human assistance, or Fin successfully completes a workflow.

This is just a change of price tag; charging per seat, per task, or per successfully resolved issue changes the unit of pricing seen by users. The application company still needs to call the model behind the scenes and calculate how many tokens are consumed to complete a delivery.

The Southern Song painter Li Song painted a piece called "Market Vendor Playing with Children." The vendor carries a load of goods through villages, with pots, bowls, needles, threads, snacks, and toys all hanging from the load. A single load is like a mobile convenience store, and the vendor earns money by selecting, preparing, and delivering goods.

AI application companies do something similar. Model manufacturers produce intelligence, and application companies bring that intelligence into specific scenarios. What is bought at the base level is tokens, and what is sold in the end should be a report, a design, or a video.

Model companies sell intelligence, while application companies sell completion.

The profit of this business mainly depends on three price differences.

The first is the procurement price difference.

For the same model, different teams may not have the same actual costs. Companies with high usage can negotiate discounts, and teams with strong engineering capabilities can use caching, asynchronous processing, peak and valley calls, and model routing to further reduce average costs. What others can complete for one yuan, you can do for fifty cents, and the remaining fifty cents is profit.

To put it simply, for the same procurement, some get wholesale prices while others get tourist prices.

The second is the usage price difference.

Some users buy memberships and then exploit AI like workhorses, while others start with high ambitions but stop using the product after half a month. As long as the average consumption of the entire user group is lower than the estimates made during the package design, the unused quota will create profit space.

This is similar to the logic of gym annual memberships and mobile data packages.

However, there is an important issue to note here. If upstream procurement is flexible enough, less user usage will increase profits. If a company has already signed a fixed annual contract, the remaining tokens lying in the account are not profit; they are just unsold inventory.

Unused tokens are called deposits, while unsold tokens are called excess inventory.

The third is the product price difference.

For the same consumption of ten thousand tokens, some products only output text that requires the user to modify for a long time; others have already completed data organization, content generation, fact verification, format adjustment, and final export. The underlying costs may be similar, but the prices users are willing to pay can be completely different.

The true value of an AI application depends on this. Procurement determines whether the business can exist, while the product determines how much the business is worth.

Reselling tokens is just the foundation; the product is the reason for the price difference.

Product Managers Become Accountants First

The trouble is that AI applications require constant procurement, so product issues can easily become cost issues.

In the official user guide for Manus, you can already see how these costs shape the product. Manus deducts points from monthly subscriptions based on the complexity of tasks and the resources consumed. When asking AI questions, the official suggests using Chat mode for simple questions and not starting a full autonomous agent; for complex tasks, it’s best to check intermediate results to avoid the agent continuing to execute in the wrong direction, wasting more points.

These suggestions are certainly reasonable. However, they also reveal one of the most troublesome aspects of agent products: the more proactive the model, the harder it is for product companies to predict costs.

When a chatbot answers once and stops, the accounting is relatively easy. An agent will split tasks, search repeatedly, open web pages, call tools, and then decide the next step based on intermediate results. The user only issues one command, but there may have been dozens of model calls behind the scenes.

As a result, products start teaching users how to make agents work less. The fear is that after calculating for too long, the accountant will start taking over the product. Whether a feature should be launched is no longer primarily based on user needs but rather on how many cents a single task will cost. The model has stronger capabilities, yet applications are hesitant to fully hand over these capabilities to users.

Companies certainly need to control costs. No startup can afford to let users use the most expensive models as if they were tap water.

Some products shove better-performing models into high-priced packages, while ordinary users receive a version that has been throttled at multiple levels. The model's capabilities do not disappear; they are just allocated to different people after the accounting is done.

Over time, the industry has developed a rather awkward inversion.

Users purchase AI applications hoping for the model to understand a bit more, think a step further, and complete more work. Application companies, however, to guard their costs, must make the model read less, think less, and do less.

The product sells intelligence, yet the team's most practiced action has become to cut back on intelligence.

Moreover, when upstream prices rise, application companies must reduce quotas; when competitors obtain lower costs, membership fees in the market will immediately drop; when new cheaper models emerge, the original routing and packages must be recalculated. Application companies seem to control the product, but the most important variable has always been in the hands of upstream.

In the past two years, AI application entrepreneurs have not necessarily lacked ideas. What they lack more is a sufficiently strong and cheap intelligence that can be confidently integrated into products.

DeepSeek Cuts Down the "Intelligence Toll"

The value of DeepSeek lies primarily in the fact that it is not a model designed solely for price reduction. It has 1 million contexts and supports reasoning modes and tool calls.

When the preview version was released in April, DeepSeek claimed its reasoning capabilities were close to V4-Pro, performing similarly on simple agent tasks. The official version updated on July 31 did not change the model structure and scale but mainly re-trained the model, with multiple agent benchmark scores already surpassing the V4-Pro preview version.

This is very important.

Cheap models have existed before. The problem is that while the cost is low, the quality cannot always be subpar. What AI applications truly need is stable intelligence at a unit price as low as possible.

At least from the benchmarks released by DeepSeek for its Agent and code, V4-Flash has crossed the "usable" line for a batch of high-frequency tasks, with prices low enough for teams to call upon frequently.

For most ordinary users, large models have likely crossed a critical point.

Ordinary people do not care about benchmarks. They rarely pay attention to whether a model leads by three or five points in a certain test; they only care about whether they can get their tasks done. When a model can read materials, make inferences, adjust tools, and deliver results, continuing to enhance its capabilities is certainly valuable. However, whether this added value can justify several times the price needs to be recalculated.

In the past, model companies claimed superiority, and higher prices seemed justified. But now, this reasoning does not hold up; we need to discuss where the costs come from, what problems the extra capabilities solve, and how many users truly need them.

DeepSeek has not made all models cheaper; it has made high-priced models something that needs to be explained to users.

Pinduoduo's early breakthrough was not just in product pricing. It showed consumers for the first time that many daily necessities could be sold at such low prices. The brand premiums, channel costs, and intermediaries that were once taken for granted suddenly needed to justify their existence.

Pinduoduo broke through the mental barrier of product pricing.

DeepSeek is breaking through the mental barrier of intelligence pricing.

Entrepreneurs Can Finally Focus on Product Development

For the AI application industry, the most important thing moving forward is that entrepreneurs can finally shift their focus back from upstream.

In the past, large models updated too quickly, making it easy for application companies to be led by them. Today, they integrate longer contexts, tomorrow they switch to new inference models, and the day after they feature Agent capabilities on their homepage. Products seem to be continuously upgrading, but the direction is not set by themselves. Whatever model vendors release, application companies showcase.

This is a typical case of upstream prosperity and downstream distraction.

Now, foundational models have the opportunity to become relatively stable production materials. Teams no longer need to revolve around model rankings every day, nor do they have to treat "integrating the latest model" as the most important product advancement.

Once models stabilize, products finally have no excuses left.

A model being able to complete a task is just the starting point. What products need to do is integrate tasks into real processes, handling permissions, data, collaboration, review, and failures, so that users do not need to teach AI how to work every time. These aspects may not be as appealing as model parameters and are hard to feature in benchmarks. However, whether users are willing to continue paying in the second month often depends on these details.

The value of AI applications lies not in presenting models to users but in making models disappear into the workflow.

When users no longer need to choose models, research prompts, or care about how many rounds of inference are invoked behind a task, applications have truly completed their work.

At this point, competition among application companies is no longer about who gets the model first; it’s about who understands a job better and who can turn the unspoken rules of the industry into part of the product.

Models can become stronger overnight, but industry understanding cannot. APIs can be accessed by everyone, but a work method that is truly accepted by users is hard to replicate.

This also gives many businesses that previously seemed too small a renewed chance to exist.

AI entrepreneurship does not necessarily have to start from the "next-generation entry point," nor does every company need to create a super Agent that serves all of humanity. It can simply solve a problem in foreign trade inquiries, serve a set of processes for chain stores, or save a segment of repetitive labor for certain professionals.

The market does not need to be large enough to accommodate everyone. As long as the problem is specific enough, the delivery is stable enough, and the value received by users exceeds the price paid, this business can survive.

In the past, when models were too expensive, small markets struggled to support a complete product. Teams had to pursue larger user scales, higher average transaction values, and faster financing speeds to cover the ever-increasing inference costs. Many entrepreneurial projects were forced to tell an overly grand story before they even found real users.

Now, the story can be a bit smaller.

Small teams can first serve a small group of people, thoroughly address one task, and then gradually expand outward.

This resembles how many "century-old stores" used to operate. When they opened, they did not think about dominating a street; they just focused on serving a few tables well at the entrance. If the food is good and the prices are reasonable, customers will return, and business will naturally thrive.

The large model industry likes to talk about scale, parameters, and endgames. Ultimately, application entrepreneurship must return to this simple logic.

Who do you serve, what problems do you solve for them, and why can’t they go to another place? These three questions, DeepSeek will not answer for entrepreneurs.

It simply allows teams to spend less time bargaining with upstream and gives small teams that lack the ability to train large models or the qualifications to access special prices a relatively fair entry ticket.

In the past, the model itself was a barrier. Moving forward, models can be purchased by anyone; the distinction lies in what they are used for. What is truly scarce will gradually shift towards users, data, processes, and trust. Those who are closer to real work will have more opportunities.

After DeepSeek lowers the threshold for intelligence, AI applications finally shift from a model competition back to a product business.

Chao Cuo from the Western Han Dynasty wrote in "On the Value of Grain": "Value the five grains and despise gold and jade."

In recent years, large models have resembled gold and jade. Parameters have grown larger, rankings have risen higher, and they are showcased under the lights at press conferences, where everyone can see their preciousness, yet very few can truly integrate them into daily work.

What the application layer needs to do is turn gold and jade into five grains.

DeepSeek has not cooked the meal for entrepreneurs; it has merely lowered the price of rice.

As for who can open the restaurant, it depends on each one's skills.

**-**END-

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.