Image source: Generated by Unbounded AI

" OpenAI CPO talked about us not worrying about the commercialization of models, because the technical curve is steep, and leaders are always defining new possibilities. Anthropic CPO admits that models are being productized, and the differences in model will become larger and larger in the future. But from the perspective of product output, they are all releasing agent products."🔥 The red sea of ​​AI Agent is coming, and giants are entering the market to compete for the beach. Hello everyone, I am Master AILin! Last week I just attended an AI entrepreneur party, and the hottest topic on site was: "If OpenAI and Anthropic enter the Agent market in full, will we start-ups still have a way out?"

X platform @StlightLeon's comment

To be honest, this question is really heartbreaking. Thinking about Manus, who was so popular some time ago, there were constant controversy, and countless imitators appeared overnight. But none of these are fatal. After all, there is still a trial application of 2 million yuan in the system! What really makes entrepreneurs restless is the sharp question: "It's just a shell, without the ability to core model."

This question really points to the core problem: If the model manufacturer does its own agent, what are the advantages of the startup company? The recent emergence of OpenAI's DeepResearch and Anthropic's Claude Code seem to confirm Alexander Doria's "model itself is a product" theory. He believes that the protagonist of the next AI development cycle is not an agent or inference system, but a specially trained model itself.

Why does this trend occur? Based on the latest technological developments, I have summarized three key reasons:

📉 General model expansion has encountered a bottleneck: OpenAI's GPT-4.5 revealed that the model capability has only increased linearly, but the computing power cost has increased exponentially. Even if training efficiency continues to improve, the cost is too high, making it difficult to deploy the latest big models widely. 📈 The reinforcement training effect of specific tasks is far beyond expectations: after combining reinforcement learning and reasoning ability training, the model shows amazing professional abilities. Small models perform well in the field of mathematics, code models can manage the entire code base, and Claude can even play games with very little information. This specialized training creates a completely new type of model. 💰 The reasoning cost has dropped sharply, and the business model is facing transformation: Taking DeepSeek as an example, the new optimization technology has enabled the existing GPU computing power to support each person to use 10,000 high-quality tokens per day. This means that the model of simply selling API calls is difficult to sustain, and model manufacturers must transform toward higher added value.

This trend not only affects the Agent's technology implementation path, but more importantly, it redefines the boundary between model manufacturers and AI application companies. Today, I will discuss a key question through in-depth analysis of the interviews between OpenAI and Anthropic CPOs: Under the new pattern of "model is a product", how can startups find breakthroughs in the Agent market?

🥊 Giant showdown: OpenAI vs Anthropic, two completely different strategies

OpenAI vs Anthropic

I spent the whole weekend carefully studying the interview content of Anthropic's Mike Krieger and OpenAI's Kevin Weil, and found that the two companies are very different in strategy. These differences may be the key to startups finding breakthroughs!

🤝 Anthropic: We are your AI partner, not just a model supplier.

Mike Krieger's interview revealed Anthropic's core strategy: to become a long-term AI partner for customers, not a simple model supplier. He emphasized:

Anthropic's big turn: no longer just make models! Bet on one side of the application and win the second half of AI | CPO Krieger 20VC reveals the new strategic focus

"We want to be your long-term AI partner to help you design products with the application AI team and imagine the future with you, not only considering APIs, but also taking overall solutions."

Krieger particularly emphasized the strategic value of first-party products (such as Claude Code), believing that these products are the key to accelerating learning, brand building and building a moat.

Interestingly, Krieger repeatedly mentions the "Day One" mentality and long-termism, showing a future-oriented patience and determination. In terms of product iteration, Anthropic seeks to find a balance between rapid innovation and stability and reliability.

What surprised me the most was that Krieger proposed the concept of "Model Identity", believing that in the future, users will not only look at functions, but also "Vibes" when choosing AI! Just like when we choose friends, we will be attracted by the "personality" of a certain AI.

⚡ OpenAI: Speed ​​is king, technology is everything.

In contrast, Kevin Weil's interview is full of Silicon Valley-style worship of speed:

"In most cases, we prioritize speed, and we prioritize putting these new tools into people's hands so that they can do more cool things faster."

Weil firmly opposes the view that "models are being commoditized" and believes that models are developing too fast, and even a 3-6-month lead is crucial. He stressed that OpenAI must become a "world-class research company and a world-class product company" at the same time.

OpenAI's mission is to make AI benefit all mankind, and to popularize AI technology as widely as possible through products and APIs such as ChatGPT and Deep Research. Weil even boldly predicts that AI will surpass humans in programming within this year, which will bring about democratization of software creation.

📊 Eight differences between the two giants

Through comparative analysis, I summarized the eight key differences in strategy between OpenAI and Anthropic:

Let's analyze the differences between these two AI giants in eight key dimensions:

1. Attitude toward the commodification of models

OpenAI firmly opposes the commodification of models, Kevin Weil stressed: "In the field of AI, technological progress is amazing. Even a 3-6-month leading advantage can make you the first company to launch revolutionary functions. Smart costs drop 10 times a year, far exceeding Moore's Law. This rapid evolution will continue to reshape the competitive landscape. We must maintain the pace of innovation and consolidate the technological leadership." Anthropic adopts a more pragmatic attitude. Mike Krieger acknowledges that there is indeed a trend of commodification at the basic model level, but he points out that true differentiation will come from three core dimensions: one is to continue to conduct breakthrough research with top AI talents, the second is to create unique model characteristics (especially in terms of security and controllability), and the third is to establish in-depth strategic partnerships with customers. He stressed: "As time goes by, the differences between models will become larger and larger, rather than convergence."

2. Data and Model Training Methods

OpenAI adopts a "teachical" training method, emphasizing the guidance of model learning through carefully designed prompts and human feedback. "The bottleneck of the model is not the intellectual limit, but how to effectively 'teach' it to accomplish tasks. We invest a lot of resources to optimize," Kevin said.Measure metrics and training environments to ensure that models can accurately understand and perform human intent. "Anthropic adopts a hybrid training strategy," Mike explained. "We first use high-quality human data to build basic cognition, and then create synthetic environments that simulate real-world complexity, allowing models to explore and grow in it. Especially when dealing with multi-step reasoning and boundary situations, this approach allows the model to gain a closer ability to practical application scenarios. "This method focuses on training the adaptability of models in real-world complexity and uncertainty.

3. Product concept

Anthropic regards first-party applications as a strategic commanding height. Products such as Claude Code not only serve users, but also serve users, but also serve them quickly and verify. Mike emphasized: "We have a deep understanding of user needs through our own products. These insights are crucial to improving the basic model." "OpenAI seeks a balance between open APIs and owned products, Kevin said: "We must not only directly serve users through products such as ChatGPT, but also empower the developer ecosystem. This is a dynamic balancing process. "

4. Innovation and Iteration Rhythm

Anthropic adopts a relatively cautious iteration strategy. Mike explained: "In the field of AI, it is difficult to rebuild once user trust is lost. We would rather be slower than make sure every update is fully verified. "OpenAI emphasizes speed first," Kevin said, "Fast iteration is our core value." On the premise of ensuring safety, we must make innovation reach users as soon as possible. "Their "Optimize for velocity" concept runs through the entire process of product research and development.

5. Product iteration strategy

Anthropic adopts a solid and balanced iteration strategy, Mike emphasized: "We seek the best balance between rapid innovation and product stability. For core functions, we will adopt a more cautious release rhythm; for cutting-edge features, users will choose independently through the 'opt-in' mechanism. This differentiated iterative strategy ensures product reliability. "OpenAI adheres to the "speed priority" concept, Kevin said: "We believe that rapid iteration is the key to promoting the development of AI." By continuously releasing new features and improvements, we can collect user feedback faster and accelerate product optimization. Even if there may be early imperfections, this speed-oriented strategy allows users to experience the latest advancements in AI more quickly. "

6. Model Differentiation Strategy

Anthropic focuses on workflow automation, Mike stressed: "OurThe goal is to create an AI assistant that can be deeply integrated into a specific workflow. Claude Code is a good example. It is not about replacing the IDE, but about improving the efficiency of the entire development process. We believe that AI in the future must not only have powerful functions, but also be able to seamlessly integrate into the way users work. "OpenAI pursues breakthroughs in general capabilities," Kevin said. "Our primary goal is to improve the basic capabilities of the model so that it can perform well in all scenarios." For example, in the field of software development, we hope that AI can not only write code, but also truly understand and solve complex engineering problems. Only this improvement of general capabilities can lead to true technological democratization. "

7. Collaborative model between research and product teams

OpenAI pursues R&D integration, Kevin emphasized: "We must be a world-class research company and product company at the same time. Research, product, engineering and design teams work in depth from the beginning of the project, rather than simple results handover. "He used Deep Research as an example to show how cross-team collaboration can drive innovation breakthroughs. Anthropic adopts a more pragmatic collaboration strategy," Mike said. "We need to find a balance between demonstrating the vision of the future and leveraging the current capabilities." The research team focuses on improving the model training environment to make it closer to the real scenario; the product team is responsible for quickly transforming the research results into available solutions. "

8. Strategic Positioning and Business Model

Anthropic adopts the "partner service" model, emphasizing in-depth customization and long-term cooperation. Mike elaborated: "Our goal is to become a strategic partner of customers, jointly design and deploy AI solutions by deeply understanding their business pain points. We provide full-process services from technical consultation, solution design to implementation operation and maintenance to help customers achieve business growth. This in-depth cooperation model allows us to continuously optimize solutions and create long-term value for our customers. "OpenAI adopts the "platform empowerment" strategy and focuses on building an inclusive AI infrastructure. Kevin emphasized: "We are committed to building an industry-leading AI platform, and through standardized APIs and development tools, developers around the world can easily access powerful AI capabilities." We believe that only by truly popularizing AI technology can we achieve the mission of "benefiting all mankind". Through products such as ChatGPT and open platforms, we are accelerating the democratization of AI technology. "

These profound strategic differences not only reflect the different visions of the two companies, but also provide startups with differentiated positioning ideas. The ideas of alienated positioning.

These differences not only reflect the different concepts of the two companies, but also imply the diversified development path of the AI ​​industry. For startups, these differences are just possible.To find a breakthrough direction!

🚀 The way to break through the startup company: differentiation, verticalization, and service

Through in-depth analysis of the interviews between the two CPOs, I found that although the model manufacturers are powerful, they also have their limitations. They are all working to bring the model closer to the post-train, but face common challenges:

Existing evaluation benchmarks cannot fully reflect the complexity of real-world tasks: Current common evaluation benchmarks (such as SWE-bench) cannot fully measure the performance of the model in real-world complex tasks. There is a need to build a more realistic training environment: it is necessary to build a more complex and realistic training environment to simulate real-world workflows, human interactions, and uncertainties. Data is key, but requires a more refined data strategy: combining human data with synthetic data, and focusing on more subtle traits such as model personality. Products and research need to be deeply integrated: it emphasizes that the product team and research team need to work more closely together to guide model training through product practice to ensure that the model can truly solve real problems. The “teaching” method and evaluation metrics are crucial: they all believe that how to “teach” the model (including data, RL environment, etc.) and how to effectively evaluate the performance of the model in real-life scenarios are the key to improving the model’s capabilities.

So, both CPOs realize that making AI models closer to real-life scenarios is a complex and challenging topic that requires continuous exploration and innovation. So, you can understand that their agent strategy is to establish a more effective feedback loop so that it can ultimately achieve the widespread application and value creation of AI technology in the real world.

These challenges provide opportunities for startups! I think startups can look for breakthroughs from the following three directions:

1️⃣ Differentiation: Don’t be stubborn and hard-working giants

Last month I shared a startup that does legal AI. They did not simply call APIs to make a chatbot, but:

Technical differentiation: Based on the open source model, special optimization is carried out for legal texts Product differentiation: Design an interface that conforms to the lawyer’s workflow, rather than a simple dialog box model differentiation: adopting a hybrid service model of "AI+human audit"

What about the results? They have obtained paid contracts from multiple law firms, and these customers are also paid users of ChatGPT!

How AI leverages the legal "gold mine", the secret that only "insiders" in Silicon Valley know | Harvey's "deep water" gold digging new logic

2️⃣ Verticalization: Being a "small and beautiful" expert Agent

A friend of mine gave up the idea of ​​being a general AI assistant and focused on AI in medical image analysis. He told me:

"Do not do itA general assistant who knows how to do anything but is not good at anything is worse than an expert who is better than humans in a specific field. "

This is the essence of the vertical strategy: deepen the vertical field, build industry knowledge-how barriers, and provide expert Agent solutions.

The data, knowledge and processes in the vertical field are difficult for giants to quickly grasp, which is the advantage of startups.

3️⃣ Service: Transfer from product sales to service delivery

Recently, I interviewed a startup company that provides AI solutions to financial institutions. Their business model shines my eyes: do not sell products, only services.

They provide:

The full implementation and operation and maintenance of customized AI integration solutions support continuous model optimization and update Professional training and consultation

Conclusion is back to the starting point: "Big model is just raw materials, and the real value lies in how to integrate it into customers' business processes and solve practical problems. "

How does inference model and AI Agent improve the productivity of knowledge workers? | Hebbia CEO dialogue a16z: AI launches a "efficiency revolution" in the financial industry

💎 Gold rush in segmented tracks: Four blue ocean markets are waiting for you to explore

Through the analysis of model vendor strategies, I discovered four blue ocean markets that startups can focus on:

🔗 "Last Mile" integration and customized services

Model manufacturers provide general models, but how to implement them in the workflow of specific industries and enterprises, there is still a huge "Last Mile "Challenge.

I visited a business last month and spent 3 months trying to integrate ChatGPT into customer service systems, and ended up encountering numerous technical and process problems. This is the opportunity for startups: to provide professional AI Agent integration and customized services.

The difficulty lies in the need of a deep understanding of customers' business processes and IT systems while solving the concerns of enterprise-level customers such as data security and privacy protection. But it is these difficulties that form the moat of startups.

🛠️ Vertical Agent Platforms and Tools

I recently tested several Agent Construction Tools for specific industries and found that they are easier to use than general platforms and their functions are more in line with industry needs.

The entrepreneurial opportunity lies in building vertical Agent Platforms and Tools, such as Agent Construction Platforms and Tools for e-commerce, education, medical and other industries, workflow orchestration tools, monitoring and management tools, etc.

A founder of Education AI told me: "A general platform is like giving you a piece of cloth and needlework, and we give you a tailor-made suit. "

👥 "Human-in-the-Loop" Agent service

No matter how powerful AI is, human judgment and experience are still irreplaceable in some key decisions and complex tasks.

I have experienced a service that provides AI-assisted writing. Their model is: first draft of AI generation, professional editor review and modification, and finally deliver the finished product. This "human-machine collaboration" model not only ensures efficiency but also ensures quality.

The entrepreneurial opportunity lies in providing "Human-in-the-Loop"Agent service", combining AI's automation capabilities with the experience and judgment of human experts, and is particularly suitable for high-risk fields such as law, medical care, and finance.

🌐 Innovation Application of Agent for Emerging Scenes

Last week I experienced a meta-universe virtual tour guide based on AI. It can adjust the explanation content and route in real time according to my interests and reactions, and the experience is very novel.

AI Agent technology is constantly expanding new application scenarios, such as meta-universe, Web3, DeFi, brain-computer interfaces, wearable devices, etc. The entrepreneurial opportunity lies in exploring innovative applications of Agent for emerging scenarios.

These emerging scenarios are not yet mature, user needs and business models are not yet clear, but the pioneers have obvious advantages and deserve the attention of entrepreneurs.

🔑 Three keys to success in entrepreneurship: professional knowledge, user experience, ecological cooperation

In communication with many AI entrepreneurs, I summarized three keys to open the door to success:

🧠 Deep Know-how: Your professional knowledge is the moat

A medical AI entrepreneur told me: "OpenAI may have a stronger model, but they do not understand the daily work flow of doctors, the particularity of medical data, and the complexity of medical supervision. These are our strengths. "

In-depth understanding of the industry, scenarios and user needs, and building professional barriers is a powerful weapon for startups to fight against giants.

🌟 Extreme experience: User experience is the most powerful differentiated weapon

"We don't need to have a stronger model than OpenAI, we just need to provide users with a better experience. "A CEO of an AI startup said, "Users don't care what model you are using, they only care whether the product can solve their problems, whether it is easy to use, and whether it is trustworthy." "

In the field of AI Agent, the ultimate user experience can be built from the following aspects:

Intuitive and easy-to-use interface design: lower the threshold for user use, and make complex AI capabilities simple and easy to use. Personalized interactive experience: Provide customized according to user habits and preferences.The interactive mode and interface are reliable and stable: Ensure that the Agent can operate stably in various scenarios without inexplicable errors. Transparent and controllable decision-making process: Let users understand the working principle of the Agent and be able to intervene and control when needed.

As Anthropic's Mike Krieger emphasizes "model personality" and "a sense of atmosphere", users' choice of AI products is not only based on functions, but also on user experience and emotional connection. Startups can make efforts in this regard and create an Agent product that users can "love".

🤝 Ecological cooperation: symbiosis and win-win are the way to sustainable development

Faced with the strong entry of model manufacturers, startups should not choose to confront each other head-on, but should seek an ecological relationship of win-win cooperation. Cooperation strategy with model manufacturers

As the "importance of API business" emphasized by OpenAI's Kevin Weil, model manufacturers need a developer ecosystem to expand their influence and application scenarios. Startups can find their own position in this ecosystem and become an important bridge connecting model capabilities and end users.

The waters are big, and AI Agent startups have great potential. The competition in the field of AI Agent is indeed becoming increasingly fierce, and the entry of model manufacturers has undoubtedly intensified this competition. But as we analyzed, model manufacturers have their capabilities boundaries and strategic priorities, which leave a broad blue ocean space for startups.

Startups should not be troubled by the fear of "what will happen if the model manufacturer does it", but should focus on their own differentiated advantages: in-depth industry knowledge-how, ultimate user experience, flexible innovation capabilities, and ecological cooperation with all parties.

In the next development cycle of AI, the argument that "the model itself is a product" may make some sense, but a more accurate statement should be that "model + application scenario + user experience + service" is the complete product. Model vendors provide strong infrastructure, but turning these infrastructure into products and services that truly solve user problems still requires innovation and efforts from startups.

As Anthropic's Mike Krieger said, AI is still in the "Day One" stage, with long-term value creationCreation is far greater than short-term competition. For entrepreneurs with vision, patience and professional accumulation, opportunities in the field of AI Agent have just begun.

In this red ocean, find the right position, give full play to its advantages, and establish win-win cooperation with all parties, and start-up companies can fully open up their own blue sky. The future has come, but it has not been evenly distributed. Those entrepreneurs who can see the trends, seize opportunities, and continue to innovate will eventually ride the wind and waves in the waves of AI Agent and head to the other side of success.