Friday, August 14, 2026

In today's science and technology circle, the most lively thing is the arms race of large models. DeepSeek V4 Pro was released, Grok 4.6 followed, and Google Gemini 3.7 Flash was not far behind. All three companies were rolling out long-term agents and programming aids. But what is most worth talking about is not whose model is stronger, but the two tricks DeepSeek is playing this time: peak and valley pricing and Harness framework. Behind these two things lies the real dilemma and future trends of the commercialization of large models.

Let's talk about peak and valley pricing first. DeepSeek raised the price this time, but it implemented a strategy of half-price in its spare time, which seemed to be the power company's routine. On the surface, it is giving profits to users, but in fact it exposes two problems: first, the computing power cost of large models remains high and cannot be sustained without price increases; second, user demand is unstable, resources are tight during peak periods, and leisure time is wasted. This can relieve some operational pressure, but it can address the symptoms rather than the root cause. The real solution is to increase model efficiency, or simply cut prices directly like Grok did-SpaceX has money to burn, but DeepSeek doesn't have the confidence to do so. What does this mean for developers? This means that in the future, when using APIs, you will have to look at the K-line just like stock trading, and avoid it when you can during peak periods, or simply build a local model yourself. In the long run, peak-valley pricing will accelerate the "decentralization" of large models, and local deployment and edge computing will be more popular.

Let's talk about the Harness framework. The core logic of DeepSeek's open source Harness this time is "Model + Harness = Agent." To put it bluntly, it is to transform the big model from a tool that can chat into an agent that can work. This framework is very ambitious. It directly targets Anthropic's Claude Code and wants to gain a piece of the desktop programming market. But there is a key issue here: agent frameworks are not scarce, but application scenarios that can be truly implemented are scarce. There are countless agent frameworks on the market, but few can really solve practical problems. Whether Harness can stand out depends on two things: first, whether DeepSeek's model capabilities are strong enough, and second, whether the framework's ease of use and ecology are good enough. At present, DeepSeek's model is indeed cost-effective, but the framework ecosystem is still in its infancy, and it is hard to say whether it will attract developers. For ordinary developers, Harness is worth paying attention to, but don't rush to all in, first look at community feedback and actual cases before deciding.

Taken together, DeepSeek is trying to build a complete ecosystem from model to application. Peak and valley pricing is to survive, Hardness is to live better. But this road is not easy. The commercialization dilemma of large models is that the value of the model itself is difficult to realize directly and must rely on upper-level applications. However, upper-level applications have high development thresholds, long cycles, and face fierce competition. DeepSeek's current strategy is to walk on two legs: on the one hand, seize the model market through price wars and performance advantages, and on the other hand, attract developers to build an application ecosystem through open source frameworks. This logic may seem reasonable, but the risks are also obvious-if the application ecosystem cannot start up, no matter how cheap the model is, no matter how cheap the model is; if the model is not strong enough, no matter how good the framework is, no matter how good it is, no matter how good it is.

The larger background is that the competition for large models has shifted from the model itself to ecology and applications. OpenAI has accumulated a first-mover advantage through the GPT Store, and Anthropic has relied on Claude Code to enter the programming market. DeepSeek wants to replicate this path, but the domestic developer ecosystem and business environment are completely different from overseas. Domestic developers are more pragmatic, value cost performance, and are more easily influenced by policies and market fluctuations. Whether DeepSeek's Harness can bear fruit in China depends on whether it can solve real development pain points, rather than just providing a cool framework.

Finally, the current big model circle is like the mobile Internet in the 2010s. Everyone is fighting for territory, but the ones that can really laugh at the end must be those companies that can solve practical problems. DeepSeek's action this time is more like a gamble-a gamble on whether model capabilities and framework ecology can form a positive cycle. The result of this gamble may take a year or two to be known. For developers, the wisest strategy now is to stay on the sidelines, not be led by concepts and hot spots, and pay more attention to actual cases and tools rather than empty publicity.

Today's technological signals reveal a clear trend: large models are evolving from mere "text generators" to "agent operating systems," and behind this transformation is the entire industry's redefinition of computing power, engineering and business models. DeepSeek's peak-to-valley pricing and the release of the Harness framework are two key footnotes to this shift.

First of all, peak-valley pricing is not just a business strategy. It exposes the structural contradiction between supply and demand of computing power in the era of large models. Unlike traditional cloud services, the computing power requirements of large models are extremely volatile-enterprise users may call centrally during the day, while individual developers prefer to use it at night. DeepSeek "flattening" the demand curve through price leverage, essentially imitating power grid dispatch: when the overall supply of computing power in society cannot increase significantly in the short term, it uses economic means to optimize the utilization efficiency of existing resources. The logic behind this is that the marginal cost of large models does not increase linearly, but shows significant differences with different periods of use. What is more noteworthy is that this pricing mechanism may become an industry standard. Google's Gemini 3.7 Flash also adopts a similar initial offer strategy, and OpenAI's next move is likely to be more fine-grained dynamic pricing. This means that the cost of using large models in the future will no longer be a fixed value, but a variable that is highly related to time, region, and task type. This is both an opportunity and a challenge for small and medium-sized businesses and individual developers-they can save costs by adjusting their working hours, but also require more complex cost management tools.

However, the real revolution comes from the Harness framework. If the big model is the "brain", then the Harness is the "limbs and nervous system." DeepSeek defines it as a "model steering layer," which means that the large model is no longer an isolated API, but an agent that can be accurately controlled. This is in direct competition with Anthropic's Claude Code and SpaceXAI's Grok 4.6, but DeepSeek's open source strategy is more disruptive. Harness's "everything is plug-in" architecture is actually building an ecosystem of agents, similar to what Android means to mobile applications. This means that future AI applications will no longer be calls from a single model, but an "agent combination" composed of multiple models, tools and services. For example, a programming agent might simultaneously call DeepSeek's code generation model, GitHub's code base search tool, and a local testing framework. The greatest value of this architecture is that it extends AI's capabilities from "generation" to "execution", thereby truly achieving automation. But the risk is also obvious: When agents can invoke tools on their own, security boundaries and accountability will become blurred. If an agent calls a malicious plug-in while performing a task, who is responsible? This will be a hot topic in the legal and ethical fields in the coming year.

These two signals together point to a larger trend: the transformation of large models from "technological wonders" to "infrastructure." This has similarities to the rise of cloud computing in the 2000s, but is more complex. Cloud computing solves the on-demand allocation of computing power, while large model agents solve the on-demand allocation of "intelligence." This means that future software development will no longer focus on writing code, but on the workflow of designing agents. For example, an e-commerce website may no longer require traditional modules such as front-end, back-end, and databases. Instead, a dedicated agent is responsible for tasks such as user interaction, product recommendation, and inventory management. This change will reshape the landscape of the entire software industry. Traditional developers may need to transform into "agent architects," and those SaaS companies that rely on fixed APIs will face the risk of being replaced by agents.

From a longer-term perspective, this shift may accelerate the process of "democratization of computing power." Today's big models are still dominated by a few giants, but open source frameworks like Harness lower the threshold for agent development, allowing small and medium-sized enterprises to build complex AI systems. This is similar to the popularity of PCs in the 1980s-when computing power shifted from mainframes to personal computers, the speed of innovation across society increased significantly. However, behind the democratization of computing power is still the hidden worry of concentration of computing power. DeepSeek's peak-to-valley pricing shows that computing power resources are still scarce, and this scarcity may spawn new forms of monopoly. For example,"computing power brokers" may emerge in the future who earn price differences by scheduling computing power resources, similar to today's electricity trading market. This will be a completely new business model and an area full of uncertainty.

Finally, it is worth noting that this technological advancement may further widen the digital divide. In today's signal, developer anxiety and social news about "spending money to buy" on V2EX are actually two sides of the same problem: the uneven distribution of technology dividends. When AI agents become infrastructure, those companies and individuals who master AI tools will gain exponential productivity improvements, while those who cannot keep up may be completely marginalized. This is very similar to machines replacing manual workers during the Industrial Revolution. The difference is that this time the change is faster and has a wider impact. How to ensure social equity while advancing technology will be one of the most important issues in the next decade.

Today, in 2026, hot events in the field of science and technology seem to be centered around artificial intelligence. The release of DeepSeek's V4 Pro and Grok 4.6, Google's Gemini 3.7 Flash model, and the launch of DeepSeek Harness all signal that artificial intelligence technology is developing at an unprecedented rate. However, behind these glories, should we see potential bubbles and risks?

First of all, DeepSeek's peak-valley pricing mechanism has attracted market attention. Although this mechanism can encourage users to adjust task time based on actual usage, it may also cause price fluctuations and cause trouble to users. In addition, although the launch of DeepSeek Harness aims to connect large models and agent products, its performance and stability have yet to be verified. If DeepSeek Harness fails to achieve the expected results, it may have a negative impact on DeepSeek's overall market reputation.

At the same time, Google's Gemini 3.7 Flash model has been continuously updated in a short period of time, reflecting the increasingly fierce competition among technology giants in the field of artificial intelligence. However, this fierce competition may also lead to waste of resources and even trigger a market bubble. In history, we have witnessed the bursting of technology bubbles many times, such as the Internet bubble and the blockchain bubble. Therefore, regarding the current AI boom, we should remain rational and be alert to potential bubble risks.

In addition, the rapid development of AI technology has also brought many ethical and legal issues. For example, the man revoked his divorce in the extra-marital embryo case, which triggered discussions on issues such as privacy and reproductive rights. This shows that while enjoying the convenience brought by AI technology, we also need to pay attention to its possible negative impacts and formulate corresponding laws and regulations to regulate its development.

In the long run, the development of artificial intelligence technology will have a profound impact on society. On the one hand, AI technology is expected to improve production efficiency and promote economic growth; on the other hand, AI technology may also lead to unemployment problems and aggravate social inequality. Therefore, while promoting the development of AI technology, we need to pay attention to its impact on society and take measures to mitigate the possible negative impact.

In short, the current hot events in the field of science and technology show that artificial intelligence technology is developing at an unprecedented rate. However, in this process, we should also be alert to potential bubbles and risks, pay attention to ethical and legal issues, and take measures to mitigate the impact of AI technology on society. Only in this way can we ensure the healthy development of artificial intelligence technology and bring real benefits to human society.