Woke up too suddenly. OpenAI is going to let AI replace your apps.

AI is no longer just a tool; it is gradually becoming a new operating system.

At this year's OpenAI Dev Day 2025, CEO Sam Altman said without hesitation that ChatGPT is not just for chatting. It aims to become an all - around AI operating system capable of running apps, managing tasks, and connecting to external services.

Just two years after the first Dev Day, OpenAI's ecosystem has undergone earth - shattering changes: the number of developers has increased from 2 million to 4 million, the weekly active users of ChatGPT have soared from 100 million to 800 million, and the number of tokens processed per minute by the API has skyrocketed 20 times, reaching 6 billion.

Meanwhile, the AI world in the past two years has also been rapidly reshaped: models are constantly iterating, agents are being widely applied, and video - generation technology is sweeping across all industries.

When the outside world thought OpenAI would keep a low profile, Dev Day on October 7th once again excited the entire industry.

At the conference, Altman launched four blockbuster products at once: GPT - 5 Pro, ChatGPT App Store, AgentKit, Sora 2.

Let's start with the most exciting part - Apps SDK. In the future, developers can directly create applications within ChatGPT.

It's not like the old - fashioned plug - ins. These are real apps that can run within the ChatGPT interface, with interfaces, interactions, login capabilities, and payment functions. The entire process can be completed here.

At the conference, several well - known applications made their debut immediately: Canva, Booking, Coursera, Zillow...

Want to make a promotional poster? Just say "Make a poster for me" in ChatGPT, and it will directly @Canva, popping up a complete design interface.

Looking for a house with a yard in New York? Just say a single sentence, and it will pull up the Zillow module, showing you clear maps, prices, and floor - plan information. The whole experience doesn't require any jumping or fragmentation, as if these applications are an integral part of ChatGPT.

Next, OpenAI showed something even cooler - AgentKit.

This is a low - code/no - code tool for those who want to "build their own agents": by dragging, connecting, and setting parameters, you can build an AI Agent that can "perform tasks".

At the conference, OpenAI's experience team built an agent in just 8 minutes to provide Q&A services for the DevDay official website.

Don't think this is just a concept show. AgentKit consists of three core modules:

• Agent Builder: A visual canvas where you can drag nodes to connect task flows.
• Connector Registry: A place to "attach tools" to the agent, such as connecting to databases, calling APIs, and accessing external resources.
• ChatKit: A chat - interaction component for embedding the agent into interfaces/apps.

There is also an Evals (evaluation system) module to measure how well an agent performs. This combination of "visualization + modularization + safety guardrails" is what OpenAI uses to turn "building an agent" from a privilege of programmers into something that "everyone can do".

Next up is Codex. At this conference, OpenAI announced that it has officially entered the GA (General Availability) phase. Even more interesting is the insider information: almost all of the team's code depends on it. About 70% of the weekly code submissions are completed with the help of Codex.

In the on - site demonstration, an engineer simply said, "Create an interface to control stage lights and cameras", and Codex immediately built the front - end UI, wrote the back - end services, searched for protocol documents on GitHub, and even integrated handle control and voice commands.

Throughout the process, the engineer didn't write a single line of code.

Codex's goal is clear: natural language + intelligent recognition = "semantic programming". It understands and acts on what you say. The "code - writing" work of programmers may gradually be replaced by a process of "instruction - conversion - verification".

At the bottom layer are several blockbuster models: GPT - 5 Pro API, gpt - realtime - mini voice model, and Sora 2 video model.

• GPT - 5 Pro debuts with a 400,000 - token context window and strong reasoning ability.
• gpt - realtime - mini reduces the cost of the voice model, making voice interaction more accessible.
• Sora 2 makes video generation controllable and synchronizes audio and video, turning video from a "heavy - asset" into a "command - controllable output".

Combining these technological forces, OpenAI aims to build a closed - loop ecosystem of "entrance - platform - application - agent".

ChatGPT is transforming from "answering your questions" to "whatever you say, I'll have the back - end services do it for you". This is not just an upgrade; it's more like a declaration of war against the existing Internet landscape.

At this point, you may ask: Can such a grand ambition be realized? I (Old Fox) also have some interesting observations to share with you.

First of all, this "unified entrance" model needs to gain the trust of application developers.

Will leading applications like Canva, Coursera, and Spotify be willing to move into the ChatGPT ecosystem? Their self - built App Stores, traffic entrances, and user data are their core assets.

Becoming a sub - module within ChatGPT means giving up some independent control. The mid - to long - term interest game in this "entrance reconstruction" is the real stormy sea.

Secondly, intermediate - state tools like AgentKit are actually very likely to be marginalized.

If you don't have strong business scenarios like Microsoft and Google, and you don't want to focus on pure underlying models, can you stand firm by integrating accurate information or providing vertical solutions at this intermediate level?

There are too many mature tools and specialized manufacturers in the market. AgentKit must have a clear positioning; otherwise, it may be "sandwiched" between competitors.

Moreover, when OpenAI gets involved in building tools and ecosystems, it may conflict with its original customers who are model - calling providers. This is the dilemma of "the referee also being an athlete". Since customers are also potential competitors, a high level of transparency and institutional guarantees are needed to build trust.

In addition, regulation and privacy are also unavoidable pitfalls. Apps SDK is not currently included in the preview list in the EU. Different countries/regions have different regulatory systems, data - circulation restrictions, and privacy - protection policies, which may become constraints during the ecosystem expansion phase.

Secondly, deep integration of applications means more calls for sensitive permissions and data interactions. If the security boundaries and data isolation are not well - established, problems can easily occur.

Finally, the model competition is not over yet. Tech giants like Google, Meta, and Anthropic are also heavily investing in models and generation systems; the open - source model camp and low - cost models are also iterating rapidly in local scenarios.

Although OpenAI is currently in the limelight, it's still unknown who will maintain the dual leadership in technology and ecosystem in the future.

Even so, judging from the signals from this conference, in the second half of the AI competition, the focus is indeed shifting from "whose model is stronger" to "who can build a comfortable, low - threshold, and secure ecosystem".

When the technological dividend is waning, the experience dividend and ecosystem dividend are the keys to victory.

What does this mean for ordinary people? Simply put, in the future, when you write an article, you don't need to install many tool plug - ins; when you do design, you don't have to switch back and forth between Canva and PPT; when you look for a house, you don't need to open several real - estate apps and map apps; when you run a company, you can build your own intelligent customer service and data - analysis robots.

You can turn a creative idea and a process into a "living" agent in just a few minutes. In the future, creativity will be more important than technology.

Today's ChatGPT may only be in the embryonic stage of this future "entrance OS". The ultimate form of ChatGPT may be when it no longer just "answers your questions" but "organizes resources to do things for you