OpenAI's Codex Is Growing Up Fast—And Getting

OpenAI has been quietly shipping a ridiculous number of updates to Codex over the past few weeks, and the cumulative effect is kind of fascinating. Not because any single feature is revolutionary—most of them aren't—but because the update velocity and the sheer range of capabilities suggest something bigger than incremental improvement. Codex is trying to become an entire coding platform, not just a better autocomplete.

AICodeKing's recent walkthrough of these updates is useful not for what it celebrates, but for what it reveals about where AI coding tools are actually heading. And it's messier, more ambitious, and weirder than you might expect.

The Desktop App Gets Serious (Mostly)

The foundational shift happened with Codex's desktop app refresh, which moved the tool from "basically a wrapper around the CLI" to something that wants to be your primary coding workspace. The headline feature: an in-app browser that lets Codex open local dev servers, let you annotate what's wrong with the UI, and then fix those issues.

This matters more than it sounds. Most AI coding tools can edit files all day long, but they're flying blind when it comes to what the application actually looks like. They can change your CSS, tweak your React components, adjust your layout—but they can't verify whether the result is good, broken, or hilariously off-brand. Browser integration closes that loop.

Codex also added macOS computer control, so it can see, click, and type inside native applications. That's useful for the kinds of bugs that don't have a clean terminal solution—GUI settings, simulator workflows, app-specific configuration.

And then there are Codex Pets.

Yes, animated companions that live inside your coding environment. You can customize them, toggle a floating overlay, even use the "hatch pet" skill to generate one based on your recent projects. It sounds absurd. It kind of is absurd. But as AICodeKing points out, "the useful part is that the floating overlay keeps your active Codex work visible while you're using other apps. It shows the active thread, tells you whether Codex is running, waiting for input, or ready for review."

So: serious enterprise permission controls and little animated pets, shipping in the same update cycle. Very OpenAI.

The CLI Updates That Actually Matter

The less flashy news is in the command-line interface, where OpenAI has been pushing out version after version—0.122.0, 0.123.0, 0.124.0, 0.125.0—each one adding foundational capabilities that won't make a good thumbnail but will determine whether teams can actually use this thing.

Version 0.122.0 introduced side conversations (ask a quick question without derailing your main thread), queued input (send commands while Codex is still working), and improved plan mode (start implementation with a clean context after messy planning discussions). None of these are exciting. All of them are the difference between a tool that's cool in demos and one that works in real workflows.

Plugins got a major upgrade too. Tabbed browsing, inline toggles, marketplace support for remote sources. AICodeKing notes this is "one of the most important directions for Codex" because "if plugins become easy enough, then Codex can become way more customizable than most closed AI coding tools." That's the right read. GitHub Copilot is locked down. Cursor is opinionated. If Codex can build a genuine plugin ecosystem, that's a meaningful differentiator.

Then there's the security layer. Deny-read policies, trusted workspace requirements, better sandbox enforcement. Boring stuff that becomes critical the moment you're working in a repo with API keys, client data, or anything remotely sensitive. This is Codex trying to grow up from "cool coding assistant" into something enterprises might actually trust.

Amazon Bedrock and the Model Wars

Version 0.123.0 added native Amazon Bedrock support, which is more significant than it sounds. Most AI coding tools lock you into specific model providers—usually OpenAI, sometimes Anthropic. Bedrock support means Codex can tap into AWS's model marketplace, which matters if you're in a company that's already deep in the AWS ecosystem and has compliance requirements around where compute happens.

And then GPT-5.5 became the recommended model for most Codex tasks, with GPT-5.4 as the fallback during rollout. The model upgrade probably matters less than the infrastructure around it—quick reasoning controls (Alt+comma to lower reasoning, Alt+period to raise it), multi-environment support, better MCP debugging tools.

AI Reviewing AI (And Why That's Not Crazy)

One of the stranger additions: automatic approval reviews. Before Codex executes certain commands, it can route them through a separate AI agent that evaluates risk and either approves, denies, or flags them for human review.

"I think this is one of those features that sounds weird at first because it is basically an AI reviewing another AI," says AICodeKing. "But for dangerous commands, file access, or risky operations, it makes sense. It gives you another layer between 'the agent wants to do this' and 'it just happened.'"

That's the right framing. It's not about perfect safety—no layer of AI review is going to catch everything. It's about reducing the blast radius when things go wrong. If you're already trusting an AI to write code, having a second AI sanity-check destructive operations before they run is not meaningfully more risky. It's just another guardrail.

What's Actually Happening Here

The pattern across all these updates is consistent: OpenAI is trying to build a complete development environment, not just a better autocomplete tool. Browser control, computer use, plugin ecosystems, multi-environment support, approval workflows, artifact viewers for PDFs and presentations.

AICodeKing puts it plainly: "Codex is not just trying to be another Claude code clone. It is trying to be a full coding workspace with browser testing, computer use, plugins, automations, PR review, artifacts, remote environments, and multiple interfaces all connected together."

Whether that vision actually works is still an open question. A lot of these features are rough. The UI has "some weird quirks," as AICodeKing diplomatically notes. And shipping features fast doesn't mean shipping them well.

But the velocity is real. Five significant CLI versions in a matter of weeks, each one adding capabilities that range from essential (better permission controls) to experimental (AI-reviewed approvals) to just plain odd (animated pets).

If you're already paying for ChatGPT, Codex is increasingly hard to ignore. If you're on a team evaluating AI coding tools, the plugin marketplace, Bedrock support, and permission system might matter more than the model picker. And if you're just watching where this technology is heading, Codex is one of the better case studies in what happens when a company decides to move very fast on an idea that's still half-baked.

The question isn't whether Codex will have bugs, rough edges, and features that don't quite work yet. It obviously will. The question is whether OpenAI can ship and iterate faster than the problems compound. So far, they're trying.

—Marcus Chen-Ramirez, Senior Technology Correspondent