Codex's recent update frequency has been insane.
Over the past two months, OpenAI has added something new to Codex almost every few days.
First came plugins, built-in browsers, computer operations, PR reviews, remote SSH, and mobile access... Then on May 21, Codex joined "Crazy Thursday," rolling out several major updates at once: one-click screen sharing with Codex, enabling Codex to focus on a goal for extended periods, continued remote usage after screen lock, and support for team-shared plugins and usage data tracking.
Previously, there was a widely circulated meme: Wake up, and see another Claude update. Now, Codex is no different.
However, Claude has been updated more frequently and with finer improvements, while Codex has released more major features.
Notably, they are both updated in the same direction—enterprise portals and real-world workflows.
Claude Code has already demonstrated the value of this path, and Anthropic has even begun convincing the market that cutting-edge model companies don’t have to burn money indefinitely—they also have the potential to achieve profitability.
Codex is doing the same thing, backed by OpenAI, which is preparing for its IPO at this time.
ChatGPT has demonstrated that OpenAI has users, but users don’t equal business, and popularity doesn’t necessarily translate to profit. Especially for a cutting-edge model company, the costs of compute power, training investment, and inference expenses are substantial. OpenAI needs to prove to the market that it’s not just capable of creating viral chatbots, but can also integrate AI into production workflows that enterprises are genuinely willing to pay for.
Codex's frequent updates are filling in this gap.
It's not just a development tool—it's OpenAI's clearest current path to demonstrating commercial value.
01
What has Codex been doing over the past two months?
We created an image using ChatGPT Images 2.0 to show what updates Codex has made over the past two months.
On March 24, search and settings are synchronized.
The Codex App now includes historical thread search and quick navigation to recent threads, while synchronizing key settings between the Codex App and the VS Code extension. This is a foundational experience improvement designed to help users quickly resume previous tasks and ensure a more consistent experience across the desktop and editor.
On March 25, the plugin system was launched.
Codex now supports plugins. Plugins can package skills, application integrations, and MCP server configurations to reuse workflows, and are supported by Codex App, CLI, and IDE extensions.
On April 9, the code review workflow was enhanced.
The Codex App has added collapsible inline review comments, different review modes, Git summaries, and source blocks. Codex is now diving deeper into code review and PR collaboration.
On April 12, file and terminal context enhancements were implemented.
Codex has added file search to the command menu, supports previewing images, PDFs, and Markdown in the sidebar, introduces a terminal tab for each thread, and allows users to directly ask Codex questions after selecting text.
April 16: Codex for almost everything.
This is the first major milestone in the past two months: OpenAI is transforming Codex into a more comprehensive AI workstation. This update includes an integrated browser, computer operations, thread automation, a task sidebar, PR workflows, result previews, SSH remote connections, multi-terminal and multi-window support, Intel Mac compatibility, and a suite of new plugins.
April 23, automated approval review.
Codex can first route eligible approval requests to an automated review agent to assess risk, then display the review status and risk level, allowing the user to decide whether to approve.
On May 5, the Codex access token was launched.
Workspace owners and administrators for 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 increasingly integrating with CI, automation, and enterprise engineering systems.
On May 7, Codex entered Chrome.
Codex has launched a Chrome extension that enables parallel work within browser tabs without directly taking over the user’s browser. Users can also control which websites are permitted to use Codex. As browsers serve as the entry point for many backend systems, internal tools, and web debugging scenarios, this step brings Codex closer to real-world office environments.
On May 14, Codex added mobile control support.
OpenAI enables users to use Codex through the ChatGPT mobile app by connecting to a Mac running the Codex app; users can also view task progress, approve actions, inspect code diffs, and see test results on their mobile devices. This update also includes the general availability of Hooks, access tokens, and enterprise admin setup guides. Codex is beginning to function as a remote workflow agent.
May 21: Appshots, Target Mode, Lock Screen Remote Access, and Plugin Sharing.
This is the second major feature. Appshots can send screenshots of the current Mac window and available text directly to Codex; the Target Mode has officially launched, allowing users to give Codex a goal and have it work continuously on that goal for hours or even days; and Lock Screen Remote Usage enables Codex to continue operating desktop applications even after the Mac is locked, eliminating the need to "keep a line open."
At the same time, ChatGPT Business now supports team-shared plugins; the built-in browser's annotation capabilities have been further enhanced to allow direct adjustments to font, color, spacing, and other styles.
The features themselves are certainly important, but the overall update trends are equally worth noting. Whether it’s Appshots, target modes, Chrome extensions, access tokens, or plugin sharing, all of these are addressing the fundamental requirements for entering real-world workflows: seeing the situation in real time, driving tasks forward, and managing risks effectively.
To see the full picture, what needs to be completed is contextual understanding.
Real development tasks rarely occur solely within a code editor. File search, file preview, terminal tabs, built-in browsers, browser annotations, Chrome extensions, and Appshots all fundamentally reduce the cost for users to describe context to AI.
Previously, you had to tell AI where the problem was by describing it or using Ctrl+C/V; now, OpenAI wants Codex to see these things directly.
To move tasks forward, long-running tasks and remote execution capabilities are essential.
The target mode addresses the question of “whether it can be sustained.” Remote access on mobile devices and remote usage while locked allow tasks to continue progressing even when users are away from their computers. Access tokens and hooks further integrate Codex into enterprise engineering systems such as scripts, schedulers, and CI runners.
Managing risk is up to the company and its team.
For individual developers, tools are judged primarily on usability, but enterprise tools involve far more complex issues: how to manage permissions, how to distribute plugins, who is using them and how much, how to conduct risk reviews, whether they can integrate with CI, and whether they can be centrally managed by the team.
Codex has also done extensive work in this area. The plugin system enables workflows to be packaged and reused; plugin sharing allows teams to uniformly distribute tools; automated approval reviews mitigate risks associated with agent execution; and access tokens and enterprise admin settings integrate Codex with existing engineering and governance processes within the enterprise.
02
The hope of the entire village
The update to Codex has brought it very impressive user growth.
In early March, Codex had around 1.6 million weekly active users. By May 14, OpenAI officially mentioned that over 4 million people were using Codex each week. This means that within roughly two months, Codex’s weekly active users more than doubled.
This growth trajectory cannot be separated from the underlying model's capabilities; users are willing to entrust Codex with real tasks more frequently only if it can actually get the job done. Especially after GPT-5.5, Codex has gained a stronger foundation in coding, tool usage, long-context handling, and multi-step task execution.
But having a model alone isn't enough—the market won't pay just because a model's benchmark improves; it cares more about whether these capabilities can generate revenue.
This is also something OpenAI must clarify before going public.
OpenAI holds many cards, but each card comes with its own uncertainty.
ChatGPT is the largest user gateway, demonstrating OpenAI's ability to attract global users and consumer-grade subscriptions. However, the larger the user base, the greater the inference costs; whether consumer-grade subscriptions can sustain long-term investment by a cutting-edge model company still needs further validation.
API is a fundamental revenue source, enabling the sale of model capabilities to developers and enterprises. However, the API market is highly susceptible to price competition, and enterprise customers may not be locked into a single model provider. The more general-purpose the model capability, the more likely customers are to use multiple models in combination.
ChatGPT Enterprise, Agents, and industry solutions represent OpenAI's direct entry into the enterprise market. However, for these products to truly integrate into enterprise workflows, they require time, sales efforts, integration, and industry-specific deployment.
Further out, OpenAI also has hardware, data centers, multi-cloud partnerships, and computing infrastructure. These stories are highly imaginative but also heavier, farther removed, and more capital-intensive. They support the long-term vision but are difficult to immediately tie to short-term business returns.
The commercial value of Codex is easier to explain, as its target audience is clearly defined: developers and engineering teams.
These are people who are already willing to pay for services. Engineers' time is expensive, software projects take a long time, and code maintenance carries high costs. Every step—bug fixes, testing, code reviews—can be quantified in terms of cost.
Software development is also one of the most critical production processes for any business. Financial companies rely on risk management and trading systems, retail companies on supply chain and membership systems, healthcare companies on data and compliance systems, and media companies on content management and distribution platforms. Even non-tech companies require extensive internal tools, data pipelines, automation scripts, and business systems to operate... Today, virtually every company depends on software systems.
In other words, Codex integrates into the areas where businesses spend money and consume human resources every day.
In a sense, it is OpenAI’s hope for crafting a compelling IPO narrative. This has become especially important as OpenAI prepares to enter the capital markets.
In the listing narrative, OpenAI is no longer facing questions like “Does AI have a future?” The real challenge is this: Can a cutting-edge model company find a clear, stable, and sufficiently profitable business path beyond massive computational investments?
More troubling is that Anthropic has already taken a step further on this issue.
03
Anthropic has taken the lead.
Codex must be brought to the forefront for another critical reason: Anthropic, one of OpenAI’s biggest competitors, has already forged a path in the enterprise space.
Although OpenAI still leads in terms of revenue, according to The Information, OpenAI’s revenue in the first quarter of 2026 was approximately $5.7 billion, higher than Anthropic’s $4.8 billion during the same period. But the issue now is no longer just how much revenue they generate—the real pressure on frontier model companies is whether their revenue growth can outpace their cost growth.
OpenAI's first-quarter revenue was high, but its adjusted operating margin was approximately -122%. On this basis, for every dollar of revenue, adjusted operating costs may amount to about $2.22, resulting in a loss of $1.22.
Over the past few years, outsiders have questioned whether large model companies are burning through too much cash—training, inference, GPUs, and talent costs all represent bottomless pits. The more users there are, the more frequent the calls, and the heavier the costs become.
Anthropic's recent signals have expanded the possibilities surrounding this matter.
According to The Wall Street Journal, Anthropic expects revenue to exceed $10.9 billion in the second quarter of 2026 and to approach its first quarterly operating profit, with an estimated operating profit of approximately $559 million.
While this doesn’t mean Anthropic has completely solved its cash-burning problem, it sends a crucial signal to the market: frontier model companies don’t have to rely indefinitely on funding—once their models are powerful enough and their products closely align with high-value enterprise use cases, revenue growth can outpace costs.
Anthropic doesn’t have a mass-market entry point like ChatGPT, nor does it pursue as many concurrent narratives. Its path is narrower and more focused: directly targeting high-value use cases where enterprises are willing to pay—particularly in areas such as developers, finance, law, research, data analysis, and internal knowledge work.
Claude Code is one of the most representative examples. It originally emerged as a must-have tool for developers, focused on programming scenarios, and gradually expanded to include tasks, plugins, permissions, team management, and enterprise governance, becoming a key entry point for Anthropic into enterprise workflows. Developers adopt it first, teams follow suit, and eventually it becomes part of corporate procurement and budgeting.
In April 2026, Anthropic’s adoption rate among Ramp’s sample enterprises rose to 34.4%, while OpenAI’s dropped to 32.3%. Although this data is based solely on enterprise spending patterns on the Ramp platform and does not represent a full market survey, it at least suggests that Anthropic is gaining momentum in enterprise paid use cases.
The pressure on Codex lies here.
OpenAI's revenue scale still leads the industry, but if it aims to enter the capital market, it cannot rely solely on user base or model capabilities. It needs a product closer to enterprise production environments to demonstrate that it can turn AI into stable corporate revenue.
If Claude Code proves that developer workflows can serve as Anthropic’s enterprise entry point, then Codex must prove that OpenAI can also succeed on this path.
Codex CEO Tibo Sottiaux recently half-jokingly summarized the company’s “master plan”: release better, more efficient models, launch improved products weekly, acquire more compute power, and spend more time scrolling on X.
A better model determines whether Codex can truly get work done; more frequent product updates determine whether Codex can integrate into real workflows; more computing power determines whether all of this can support increasingly large usage.
These are all crucial for listing.
In other words, Codex’s recent flurry of updates isn’t just about catching up on features—it’s also about following the enterprise path that Anthropic has already forged.
ChatGPT has demonstrated that OpenAI has users.
And Codex aims to prove that OpenAI is a profitable business.
This article is from the WeChat public account "Letter AI," authored by Yuan Xinyue.