The revived Codex has shouldered the hope for OpenAI's IPO.

The recent update frequency of Codex is truly insane.

In the past two months, OpenAI has almost stuffed a new feature into Codex every few days.

First came plugins, built - in browsers, computer operations, PR reviews, remote SSH, mobile access... Then on May 21st, Codex also had its own "Crazy Thursday", updating several major features in one go: handing over the screen content to Codex with one click, allowing Codex to work around a target for an extended period, continuing remote use after the computer is locked, and supporting team - shared plugins and viewing usage data.

Previously, there was a widely circulated meme on the internet: waking up to see another update of Claude. Now, Codex is no less impressive.

It's just that Claude's updates are more "fragmented" and refined, while Codex has released more major features.

Notably, their updates are all in the same direction - the enterprise entrance and real - world workflows.

Claude Code has already proven the value of this path. Anthropic has even started to make the market believe that cutting - edge model companies don't always have to burn money and may have the opportunity to turn a profit.

Codex is also doing the same thing. At this point, behind it stands OpenAI, which is preparing for an IPO.

ChatGPT has proven that OpenAI has users, but users don't equal business, and popularity doesn't necessarily bring profit. Especially for a cutting - edge model company, the costs of computing power, training investment, and inference are all high. OpenAI needs to prove to the market that it's not just good at creating popular chatbots but can also integrate AI into the production processes that enterprises are truly willing to pay for.

Codex's high - frequency updates are filling in this gap.

It's not just a development tool; it's the most straightforward card for OpenAI to explain its commercial value at the moment.

01

What has Codex been doing in the past two months?

We used ChatGPT Images 2.0 to create a picture to see what updates Codex has made in the past two months.

March 24th: Search and settings synchronization.

The Codex App added the functions of historical thread search and quick jump to recent threads, and synchronized the key settings in the Codex App and the VS Code extension. It's basically an optimization of the basic experience: allowing users to retrieve previous tasks more quickly and making the usage experience on the desktop and in the editor more consistent.

March 25th: The plugin system went live.

Codex started to support plugins. Plugins can package skills, application integrations, and MCP server configurations to reuse workflows and support the Codex App, CLI, and IDE extension.

April 9th: Enhancement of the code review workflow.

The Codex App added collapsible in - line review comments, different review modes, Git summaries, and source blocks. Codex began to get more deeply involved in code review and PR collaboration.

April 12th: Enhancement of file and terminal context.

Codex added file search in the command menu, supported previewing images, PDFs, and Markdown in the sidebar, added a terminal tab to each thread, and allowed users to directly ask Codex after selecting text.

April 16th: Codex for almost everything.

This was the first major milestone in the past two months. OpenAI started to promote Codex as a more complete AI workbench. This wave of updates included a built - in browser, computer operations, thread automation, task sidebar, PR workflow, result preview, SSH remote connection, multi - terminal, multi - window, Intel Mac support, and a batch of new plugins.

April 23rd: Automatic approval review.

Codex can first hand over eligible approval requests to an automatic review agent to assess risks, then display the review status and risk level, and finally let users decide whether to approve.

May 5th: Codex access tokens went live.

Workspace owners and administrators of ChatGPT Enterprise can allow members to create Codex access tokens for use in trusted non - interactive local workflows such as scripts, schedulers, and private CI runners. Codex is getting closer to CI, automation, and enterprise engineering systems.

May 7th: Codex entered Chrome.

Codex launched a Chrome extension that can work in parallel in browser tabs without directly taking over the user's browser. Users can also control which websites allow Codex to be used. The browser is the entrance for many backend systems, internal tools, and web debugging scenarios. This step brings Codex closer to the real - world office environment.

May 14th: Codex supports mobile control.

OpenAI allows users to use Codex from the ChatGPT mobile app and connect to a Mac running the Codex App. Users can also check task progress, approve operations, view code differences, and test results on their mobile phones. This wave also included the official availability of Hooks, access tokens, and enterprise administrator setting guides. Codex is starting to become a work agent that can be remotely monitored.

May 21st: Appshots, target mode, remote use after locking the screen, and plugin sharing.

This was the second major milestone. Appshots can directly send screenshots and available text from the current Mac window to Codex; the target mode was officially launched, allowing users to give Codex a target and let it work around this target for hours or even days; remote use after locking the screen allows Codex to continue operating desktop applications after the Mac is locked, without the need to keep the computer awake.

Meanwhile, ChatGPT Business started to support team - shared plugins; the annotation ability of the built - in browser was further enhanced, allowing direct adjustment of styles such as font, color, and spacing.

The features themselves are of course important, but the overall update trend is also worth noting. Whether it's Appshots, target mode, Chrome extension, access tokens, or plugin sharing, they are all filling in the basic conditions for entering real - world workflows: seeing the on - site situation, driving tasks forward, and controlling risks.

To see the on - site situation, what needs to be filled in is contextual ability.

Real - world development tasks rarely occur only in code editors. File search, file preview, terminal tabs, built - in browser, browser annotation, Chrome extension, and Appshots are essentially all reducing the cost for users to describe the context to AI.

In the past, you had to tell AI where the problem was through description or Ctrl + C/V. Now, OpenAI wants Codex to directly see these things.

To drive tasks forward, long - task and remote execution capabilities are important.

The target mode addresses the issue of "whether it can keep going". Remote access from mobile devices and remote use after locking the screen allow tasks to continue even when users are not in front of the computer. Access tokens and Hooks further integrate Codex into enterprise engineering systems such as scripts, schedulers, and CI runners.

Whether risks can be controlled is a matter of enterprise and team management.

For individual developers, the core of using tools is whether they are easy to use. However, the issues of enterprise tools are much more complex: how to manage permissions, how to distribute plugins, who is using them and how much, how to review risks, whether they can be integrated into CI, and whether they can be managed uniformly by the team.

Codex has also done a lot of work in this regard. The plugin system allows workflows to be packaged and reused; plugin sharing allows teams to distribute tools uniformly; automatic approval review controls the risks of agent execution; access tokens and enterprise administrator settings integrate Codex into the existing engineering and governance processes of enterprises.

02

"The hope of the whole village"

Codex's updates have brought it a very impressive user growth rate.

At the beginning of March, the weekly active users of Codex were around 1.6 million. By May 14th, when OpenAI officially introduced the mobile version of Codex, it was mentioned that the number of people using Codex every week had exceeded 4 million. That is to say, in about two months, the weekly active users of Codex have increased significantly.

This growth cannot be separated from the capabilities of the underlying model. The premise for users to hand over real - world tasks to Codex more frequently is that it can actually do the work. Especially after GPT - 5.5, Codex has a better foundation in coding, tool invocation, long - context, and multi - step task capabilities.

However, having a good model is not enough. The market won't just pay because a model's benchmark score is high. It is more concerned about whether these capabilities can turn into revenue.

This is also something that OpenAI must clarify before going public.

OpenAI has many cards in hand, but each card has its own uncertainties.

ChatGPT is the largest user entrance, proving that OpenAI has global users and consumer - level subscription capabilities. The problem is that the larger the user scale, the heavier the inference cost; whether consumer - level subscriptions can support the long - term investment of a cutting - edge model company still needs to be proven.

API is the basic source of income, which can sell model capabilities to developers and enterprises. However, the API market is prone to price competition, and enterprise customers may not be bound to a single model provider. The more general the model capabilities, the more likely customers are to use multiple models in combination.

ChatGPT Enterprise, Agents, and industry solutions are the front - line battlefields for OpenAI to enter the enterprise market. However, for these products to truly penetrate into enterprise processes, it takes time, sales, integration, and industry implementation.

Looking further ahead, OpenAI also has hardware, data centers, multi - cloud cooperation, and computing power infrastructure. These stories have great imagination, but they are also heavier, more distant, and more money - burning. They can support the long - term vision but are difficult to explain short - term commercial returns immediately.

And the commercial value of Codex is easier to explain. Its target audience is clear: developers and engineering teams.

This is a group of people who are already willing to pay for services. Engineers' time is valuable, software project cycles are long, and code maintenance costs are high. Costs can be calculated for every link, such as bug fixing, testing, and code review.

Software development itself is also one of the most core production processes of enterprises. Financial companies have risk control and trading systems, retail companies have supply chain and membership systems, medical companies have data and compliance systems, and media companies have content back - ends and distribution systems. Even non - technology companies have a large number of internal tools, data pipelines, automation scripts, and business systems to maintain... Almost all companies today rely on software systems.

That is to say, Codex is cutting into the areas where enterprises spend money and consume manpower every day.

In a sense, it is the hope for OpenAI to tell a good IPO story. At the time when OpenAI is preparing to enter the capital market, this becomes especially important.

Because in the IPO narrative, OpenAI is no longer facing the question of "whether AI has a future". The really difficult question to answer is another thing: can a cutting - edge model company find a clear, stable, and profitable business path beyond huge computing power investment?

What's more troublesome is that Anthropic has already taken a step forward in answering this question.

03

Anthropic has taken the lead

There is a very crucial reason why Codex must be pushed to the forefront: Anthropic, one of OpenAI's biggest competitors, has already blazed a trail in the enterprise market.

Although in terms of revenue scale, OpenAI still leads. According to The Information, OpenAI's revenue in the first quarter of 2026 was about $5.7 billion, higher than Anthropic's $4.8 billion in the same period. But now the problem is not just about how much revenue there is. The real pressure on cutting - edge model companies is whether revenue growth can outpace cost growth.

OpenAI had high revenue in the first quarter, but its adjusted operating profit margin was about - 122%. Calculated by this standard, for every $1 of revenue, the adjusted operating cost may be about $2.22, resulting in a loss of $1.22.

In the past few years, the outside world has always questioned that large - model companies burn too much money: training, inference, GPU, and talent expenses are all bottomless pits. The more users there are and the more invocations there are, the heavier the cost.

The signals recently released by Anthropic have changed the imagination of this situation.

According to The Wall Street Journal, Anthropic expects its revenue in the second quarter of 2026 to exceed $10.9 billion and approach its first quarterly operating profit, with an expected operating profit of about $559 million.

Although this doesn't mean that Anthropic has completely got rid of the money - burning problem, it gives the market a very important signal: Cutting - edge model companies don't always have to rely on financing to survive. As long as the model capabilities are strong enough and the products are close enough to high - value enterprise scenarios, revenue growth may outpace cost growth.

Anthropic doesn't have a mass - market entrance like ChatGPT, nor does it have as many concurrent stories. Its path is narrower and more focused: directly entering the areas where enterprises are willing to pay, especially high - value scenarios such as developers, finance, law, research, data analysis, and internal knowledge work.

Claude Code is the most typical example. It was initially a must - have tool in the developer community, focusing on programming scenarios. Later, it gradually added long - task, plugin, permission, team management, and enterprise governance features and slowly became an important entrance for Anthropic to enter enterprise workflows. Developers start using it first, then teams follow, and finally it becomes part of enterprise procurement and budgets.

In April 2026, among the sample enterprises on Ramp, Anthropic's adoption rate rose to 34.4%, while OpenAI's dropped to 32.3%. Although this is only based on the enterprise expenditure samples on the Ramp platform and doesn't represent the full - market statistics, this data at least shows that Anthropic's momentum in enterprise - paid scenarios is getting stronger.

This is where the pressure on Codex lies.

OpenAI still leads in revenue scale, but if it wants to enter the capital market