Text | Chanlianhui CLS
OpenAI is shifting its focus from ChatGPT to Codex.
On June 2, local time, OpenAI announced that in the coming weeks, it will integrate Codex, ChatGPT, and the browser product Atlas into a desktop super application. This reveals an important underlying change: internally, OpenAI increasingly believes that what truly represents the next generation of AI may not be ChatGPT, but Codex.
Initially positioned as a programming tool, this product has now become one of OpenAI's fastest-growing new businesses. Over the past year, Codex's weekly active users have rapidly increased to over 5 million; enterprise customers continue to grow; the company has even reorganized its internal structure around Codex, allocating more resources to this direction.
The reason is straightforward.
Over the past two years, ChatGPT addressed the need for answering questions. Codex is now addressing the need for getting work done.
For enterprises, the value of the two is not the same: one merely tells you the answer, the other directly does the work.
This is why, over the past year, an increasing number of analysts, investment managers, bankers, marketers, designers, and product managers have begun pouring into Codex. OpenAI has found that in many complex tasks, Codex's performance already surpasses that of ChatGPT.
This simultaneously reflects a significant shift underway in the AI industry: the chat era might be approaching its ceiling; the execution era is just beginning.
Why Bet Heavily on Codex?
If we rewind one year, Codex was just one among many of OpenAI's products.
At that time, the industry's focus remained on chatbots. Whether it was ChatGPT, Claude, or Gemini, the essence was competing on who was smarter and who answered questions more accurately.
But soon, a change began to emerge.
In 2025, Anthropic launched Claude Code, which quickly gained popularity among developers. Compared to traditional chatbots, Claude Code could directly modify code, call tools, and execute complex tasks, demonstrating significantly higher work efficiency.
This gave OpenAI its first sense of pressure. They realized that the scenarios enterprises were truly willing to pay for were not chatting, but work.
Consequently, the company began increasing its investment in Codex.
After that, with the successive releases of models like GPT-5.2 and GPT-5.5, Codex's capabilities improved rapidly. From an initial code completion tool, it gradually evolved into an AI agent capable of autonomously calling tools, handling complex workflows, and completing long-chain tasks.
User growth also began to accelerate. Over the past few months, Codex's weekly active users grew from 3 million to 4 million, then broke through 5 million.
Enterprise revenue similarly saw rapid improvement.
A consensus gradually formed within OpenAI: Codex might not be a supplement to ChatGPT, but the core product of the next phase. Because the problems they solve are completely different.
ChatGPT is more like an advisor; you ask a question, and it gives advice. Codex is more like an employee; you give a goal, and it takes care of the execution.
Take a simple example.
If you ask ChatGPT to analyze a listed company, it will tell you about the company's situation, industry background, and potential investment logic. Codex, however, might directly read financial reports, build models, perform comparable company analysis, and ultimately output a complete research report.
The difference between the two is not the quality of the answer, but the scope of work.
This change ultimately prompted OpenAI to begin reorganizing its internal structure.
The importance of the Codex team has continued to rise; products, platforms, and toolchains have begun to be reintegrated around Codex; and the upcoming super application is essentially aimed at merging ChatGPT and Codex into a unified entry point.
Because OpenAI has become increasingly aware:
The most important competition in future AI may not be about who chats better, but about who can better complete work for users.
Codex Reshaping the Work Methods of White-Collar Positions
What truly shook the industry was not Codex surpassing 5 million users, but who is using Codex.
OpenAI's latest disclosed data shows that among new users in the past month, approximately 40% are no longer developers. Analysts, investment managers, bankers, marketers, operations personnel, product managers, designers, and researchers are becoming the fastest-growing groups for Codex.
This means Codex is transforming from a programmer's tool into a knowledge work platform.
OpenAI itself is the most typical case. Within the company, non-technical teams have already started using Codex to create executive presentation materials, build internal applications, create business analysis dashboards, and automatically transform marketing ideas into content that complies with brand guidelines.
For external clients, the situation is even more evident.
For example, employees at Zapier have begun using Codex to automatically organize information from Slack, Google Docs, and Coda, then generate project reviews, incident response plans, and product requirement documents.
NVIDIA's research team uses Codex to explore research directions, manage experimental processes, and even write machine learning infrastructure scripts.
To drive this change, OpenAI, through this product merger, directly launched six plugins targeting different professions:
A data analysis plugin for analysts and business teams; a sales plugin for salespeople; a product design plugin for product managers and designers; a creative production plugin for marketing departments; a public stock investment plugin serving investment institutions; and an investment banking plugin directly serving investment banking practitioners.
A closer look reveals that these plugins almost cover the core scenarios of traditional white-collar work.
In the past, an investment manager analyzing a listed company could take several hours or even days. Now, they just need to upload the materials, and Codex can automatically complete data extraction, financial analysis, peer comparison, and investment logic organization.
Similarly, in the past, a product manager needed to first write a requirements document and then find a designer to create a prototype. Now, Codex can directly generate an interactive interface based on the idea.
This is also the change OpenAI values most, because enterprises are willing to pay for efficiency, and execution capability is far more likely to create commercial value than chatting capability.
In a sense, Codex is becoming an automation tool for knowledge work. It may not necessarily replace the jobs themselves, but it will definitely reshape how those jobs are performed.
What OpenAI Wants to Build is Not Just an AI Tool
If the rise of Codex were merely a product story, its significance would be limited.
What truly deserves attention is that OpenAI is leveraging Codex to redefine the future form of software.
At this launch event, OpenAI introduced a new feature called "Sites." Simply put, after users upload documents, spreadsheets, financial models, or project materials, Codex can directly generate an interactive website.
Client presentations, project management, product launch centers, and financial scenario analysis models can all be turned into websites.
Work that previously required PowerPoint, Excel, and Word to accomplish separately can now be done through an AI-automatically generated and continuously updated website.
Simultaneously, OpenAI also introduced an "Annotation" feature. Users no longer need to regenerate entire content; they can directly select a part to modify, such as changing a chart, adjusting text, or updating data sources.
The entire process is increasingly similar to human-to-colleague collaboration, rather than human-to-software interaction.
Behind this change lies an even greater ambition for OpenAI.
Over the past few decades, the logic of the software industry has been: users learn software and then use it to complete work.
OpenAI is now trying to reverse this: have AI understand the work, then automatically call upon software, with users only needing to state the objective. As for which tools to call, what steps to execute, and what results to generate, all are left to the AI.
From this perspective, Codex is no longer just a code tool; it's more like a digital employee.
This also explains why OpenAI is willing to restructure its entire product system around it. Because in the future, users might not care at all whether they are using ChatGPT, Codex, or another model. They only care about one thing: whether the task was completed.
References:
"Codex for every role, tool, and workflow", OpenAI;
"OpenAI Reorganizes Around Codex as Usage Surges", The Information.
