OpenAI’s Codex Update

Last night, just as I was about to shut down my computer, I received a notification on my phone. OpenAI had published a blog titled “Codex for (almost) everything.” I clicked in and was stunned after just two minutes of reading.

This tool, which was merely a code completion assistant three months ago, now claims it can control my computer, has a built-in browser, generates images, and includes over 90 plugins!

I confirmed twice—this is not a new feature of ChatGPT; this is Codex.

Six New Capabilities: OpenAI is Not Holding Back

Thibault Sottiaux, the Head of Codex, made a crucial statement: “Codex is our most capable agent; it runs on your computer, and we are just following this path forward.” He then presented two sets of numbers: over 3 million developers use Codex weekly, with 50% of them using it for more than just coding.

Half of the users are treating it as a personal assistant. This figure is somewhat alarming.

This update adds six major capabilities: Computer Use, built-in browser, image generation, 90+ new plugins, memory, and automation. In simple terms—Codex wants to manage everything on your computer, not just help you write code.

The timing of this update is quite strategic. Just days after Claude Opus 4.7 was released, OpenAI unveiled its powerful features, clearly targeting Anthropic. An analysis on 36Kr noted that the more frequent updates become, the more similar Claude Code and Codex appear. Indeed, the product lines of these two companies are beginning to mirror each other. The landscape is evolving, and the AI programming arena resembles an arms race. However, I feel something is off.

Computer Use: A Feature Previously Introduced by Anthropic

Among the six capabilities, the most groundbreaking is undoubtedly Computer Use.

Codex can now directly operate applications on your Mac. It can click the mouse, type on the keyboard, and view the screen—all while running in the background without interfering with your cursor. You can write code in one window while it runs tests and takes screenshots in another.

An official demonstration showed Codex compiling a Tic-Tac-Toe game in Xcode, playing a round, identifying a bug in the branching logic, fixing it, and validating the changes—all without human intervention.

This is somewhat terrifying.

However, it’s important to clarify something.

Claude’s Computer Use feature was released in March this year as a research preview. OpenAI’s approach is different; they have packaged Computer Use into Codex alongside the other five capabilities.

The path divergence is clear. Claude offers “I have a Computer Use feature,” while Codex states, “Computer Use is just one-sixth of my capabilities.”

To be honest, the competition has become suffocating. Yet, there’s an awkward fact—this feature is currently only available on macOS, and users in the EU, UK, and Switzerland cannot access it yet. Windows users? Just wait.

Additionally, you need to install the Computer Use plugin and enable Screen Recording and Accessibility permissions in system settings. It’s a hassle, and the barrier to entry is not low for beginners.

The direction is clear. Future programming tools will not just assist you in writing code; they will genuinely sit beside you and operate your computer. Honestly, this thought is quite surreal and a bit chilling.

90+ Plugins: Not Just a Coding Tool, But a Personal Assistant

Computer Use is the showcase feature, but the plugin system may have the most profound impact.

Codex’s plugins are a combination of three components: skills (task descriptions), app integrations (application permissions and interfaces), and MCP servers (backend data and tool sources). This update has added over 90 plugins.

Notable mentions include Atlassian Rovo (managing JIRA), CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Superpowers, and more.

An interesting prompt from the demonstration was: “Can you check Slack, Gmail, Google Calendar, and Notion and tell me what needs my attention?” One command crosses four plugins, allowing Codex to assist with your daily standup.

This is no longer just coding. You are asking a “coding tool” to check your emails, calendar, and Slack messages—what’s the difference from hiring a secretary?

The key is that there is a priority ranking among the plugins. When you ask Codex to perform a cross-tool task, it will determine the order of execution based on the plugin library.

A struggle for ecological niches is happening here. Honestly, I feel a bit uneasy. Are you sure you want an AI reading your Slack messages, Gmail, and Notion notes simultaneously? Is that safe? Who is responsible if a bug occurs?

It’s overwhelming.

Memory + Automation: Codex Begins to Have a Sense of Time

Memory is straightforward; Codex will remember your preferences, changes you made, and the technical background you explained last time, saving you from repeating yourself.

Automations are what truly deserve attention.

Previously, automation merely meant running a script on a schedule. This time, there’s a change: automated tasks can use a previously run conversation thread, carrying over the context to continue. An Automation can schedule its next trigger days in advance and can execute across weeks.

Some users are employing it as a PR babysitter, monitoring unmerged PRs within their teams. Others use it as a customer feedback monitor, tracking feedback across Slack, Gmail, and Notion.

However, these features are currently only available to users in the U.S. Users in the EU, UK, educational, and enterprise versions will have to wait. This regional limitation is a persistent issue that OpenAI has not yet addressed.

Looking back at trends, in the past six months—Claude first introduced Computer Use, and OpenAI followed suit; Claude launched the MCP protocol, and OpenAI developed its plugin system. Both companies are moving towards a direction of “not just writing code.”

The analysis from 36Kr is correct: the homogenization of underlying large models will inevitably lead to convergence in upper-level applications. The valuations of the two companies in the secondary market are closely intertwined. Capital’s sensitivity is acute; in their eyes, two unicorns are growing similar horns.

As this battle reaches its climax, the ones suffering are the developers still hesitating whether to get on board. If you don’t join, your efficiency will lag significantly; if you do, you might feel led by AI.

Some Easily Overlooked Features

Beyond the six major capabilities, several smaller features have been added.

PR review—Codex can recognize review comments written by colleagues on GitHub PRs and process them one by one. Multi-terminal tabs—open multiple terminals in Codex to run in parallel. SSH connection to remote devbox—Codex can be attached to a remote development machine, currently in alpha stage.

The sidebar file preview is particularly emphasized. Codex is beginning to cater to a type of “product deliverable for non-developers”: running an agent to generate xlsx or pdf files that can be opened, screenshot, and sent to others.

In the future, even non-coders will be able to use Codex. Generating reports, performing data analysis, and outputting documents—tasks that previously required developer assistance can now be accomplished with a single prompt.

Wait a minute. Isn’t this bad news for programmers?

What Should Developers Do?

Here’s a possibly unpopular opinion—the competition among AI programming tools is not entirely bad for us.

As Claude Code and Codex become more similar and mutually roll out features, users benefit more. The 3 million weekly active developers may become 10 million in six months. When everyone is using AI to write code, the differentiated value will return to individuals—architectural judgment, requirement understanding, cross-domain association, which AI cannot yet replace.

But I truly have my doubts.

50% of users are already using Codex for more than just coding. When a programming tool starts managing your schedule, checking emails, and monitoring PRs, it is no longer just a tool. It permeates every aspect of your work.

Where is the boundary between convenience and control?

Perhaps we will only have answers looking back six months from now. Or perhaps there will be none.

However, one thing I can do right now is to revisit Codex’s update log to see if I missed anything. The first step to not being replaced by AI is to understand how far AI has evolved.