OpenAI’s latest Codex update basically turns it from “that coding assistant you call when you’re stuck” into an all-purpose developer sidekick that quietly runs your computer, remembers how you like to work, and keeps your projects moving even when you’re off doing something else.

If you’ve been thinking of Codex as just “ChatGPT, but for code,” this release is the moment that framing really starts to break. Codex is now a desktop environment, a background agent, a reviewer, a build-and-release helper, and a kind of project memory layer for your whole dev workflow.

At the center of this shift is a simple but powerful idea: Codex should be able to do almost anything a developer can do on a computer, not just inside an editor window. That’s why the new version leans so hard into computer control, plugins, long-running tasks, and memory, all built on top of GPT-5.3-Codex under the hood.

The headline upgrade is computer use. Codex can now literally operate your machine with its own cursor in the background, seeing your screen, clicking around, and typing into apps while you keep working in other windows. That means instead of just generating code snippets, Codex can actually open your tools, run your app, tweak a config, reload a page, or poke through a UI the same way a human teammate would. For frontend work, it’s especially interesting: you can ask it to iterate on a design, and it will use the in-app browser to load your local build, apply changes, and react to what it sees visually.

The “multiple agents” angle is where it starts to feel like a small team living inside your Mac. OpenAI says you can run several Codex agents in parallel on your desktop, all with their own cursors, without them stepping on your own work in other apps. In practice, that could look like one agent hammering on UI bugs in a dev build while another runs integration tests in a terminal and a third keeps an eye on logs or dashboards. You’re not just delegating single prompts; you’re dispatching background workers.

Codex also moves closer to your browser with an embedded in-app browser you can annotate directly. Instead of trying to describe a layout issue in words, you can literally comment on the page inside Codex and say “fix this spacing” or “match this card to our pricing page style.” Today, that’s focused on localhost web apps and game development, but OpenAI is clear about the direction of travel: over time, Codex should be able to command the browser more fully, not just for development use cases.

Another big piece is images. Codex now taps into OpenAI’s gpt-image-1.5 model to generate and iterate on visuals in the same workspace you use for code. On paper, that sounds like a nice-to-have, but in a real-world flow, it matters: sketching product concept art, quickly mocking up a UI variant, or generating in-game assets no longer requires jumping out to another tool and then trying to reconcile that with your code. For small teams, that’s the difference between “someday we’ll design that” and “let’s just have Codex draft a version and see how it feels.”

Underpinning all of this is a huge plugin expansion. OpenAI is rolling out more than 90 additional plugins that connect Codex to the tools developers live in: Atlassian Rovo for wrangling Jira, CircleCI for pipelines, CodeRabbit and GitLab Issues for reviews and tracking, Microsoft’s productivity suite, Neon by Databricks, Remotion, Render, Superpowers, and more. Instead of trying to be an all-in-one monolith, Codex leans into being a hub: it grabs context from your systems of record, takes actions inside them, and surfaces the next thing you should care about.

This all lands on top of GPT-5.3-Codex, which OpenAI pitches as its most “agentic” coding model so far. It’s designed to support the entire software lifecycle: writing and refactoring code, debugging, deploying, monitoring, writing PRDs, editing copy, and handling a lot of the glue work that sits between those activities. Benchmarks aren’t everything, but OpenAI cites new highs on SWE-Bench Pro and Terminal-Bench for coding and real-world agent behavior, with strong performance on OSWorld and GDPval for more general knowledge work, all while running about 25 percent faster than the previous Codex generation.

If you step back and look at the workflow, Codex is clearly moving from “I ask for snippets” to “I work alongside an agent for the whole lifecycle.” The app now supports addressing GitHub review comments, running multiple terminals, and even connecting to remote devboxes over SSH (currently in alpha). You can open files directly in a sidebar with rich previews for PDFs, spreadsheets, slide decks, and documents, and a new summary pane keeps track of what the agent is planning, which sources it’s using, and which artifacts it has produced. Put differently, Codex is turning into a kind of living project dashboard, not just a chat box.

Developers who have been living with Codex day to day say the improvements are not just cosmetic. In a long-running review from early 2026, one engineer describes how multi-turn conversations and branch-aware updates finally feel solid enough for complex refactors. Codex can push follow-up commits to the same branch, maintain a back-and-forth about implementation details, and respect existing code style and TypeScript types across files, which is the kind of detail work that made earlier AI tools feel brittle. Tasks that reliably failed in 2025—like migrating a complicated auth system across multiple modules—now routinely succeed, and the failure mode has shifted from “mysterious crash” to “here’s why this approach won’t work, try this instead.”

Beyond code, OpenAI is leaning hard into “carry work forward over time.” Automations now let Codex reuse existing conversation threads, so the context you built up with it yesterday or last week isn’t lost the next time you open the app. Codex can schedule future work for itself and then wake up autonomously to keep going on a long-running task, potentially over days or weeks. Teams are already using this to land open pull requests, chase down lingering todos, and watch fast-moving channels in Slack, Gmail, and Notion.

Memory is another preview feature that’s quietly a big deal. Codex can now remember personal preferences, corrections, and hard-won context from past interactions. That means you can tell it once that you prefer Tailwind and dark mode for React components or that your team uses a particular branching strategy, and it will apply that the next time by default. Combined with plugins and project context, Codex can proactively suggest “useful work” when you sit down in the morning: surfacing unresolved Google Docs comments, pulling relevant context from Slack and Notion, and turning all of that into a prioritized action list.

From a developer’s perspective, the most important question is whether this actually changes the day-to-day feel of the job. The emerging consensus is that Codex is starting to swallow the grunt work: maintenance tasks, doc fixes, dependency bumps, test coverage improvements, and the endless backlog of small-but-annoying issues that never quite make it to the top of a sprint. One engineer estimates Codex has absorbed 30–40 percent of the implementation “busywork,” with human developers focusing more on architecture, product decisions, and tricky edge cases. It doesn’t replace the hard parts of engineering, but it does make everything else more fluid.

Of course, there are tradeoffs. Letting an AI operate your desktop, read your screen, and wire into your tools raises obvious security and privacy questions. OpenAI’s documentation emphasizes that computer use only activates with explicit permission and that you can restrict what Codex sees and controls on your Mac, but it will still force teams to think harder about policies, access levels, and audits. For many organizations, especially in regulated industries, the conversation won’t just be “can this help us ship faster?” but “what are we comfortable letting an agent actually click and type on in production?”

On availability, the rollout is measured rather than explosive. The new features are starting to land for Codex desktop app users signed in with ChatGPT accounts, with computer use coming to macOS first and European regions following after. Context-aware personalization and memory features will reach Enterprise, Education, and EU/UK users a bit later. If you’ve only used Codex inside a terminal or editor extension until now, OpenAI is clearly trying to coax you into moving more of your workflow into the dedicated app—and, by extension, into this more agentic, desktop-wide model of collaboration.

What’s striking is how quickly this all happened. Codex, as a product, is barely a year old in its current form, and yet developers have already shifted from treating it as an occasionally helpful autocomplete to relying on it for PR reviews, system understanding, debugging, and ongoing project coordination. OpenAI says its broader mission is to narrow the gap between what people can imagine and what they can build, and this Codex release is very much in that vein. By moving closer to the tools, workflows, and decisions that define real-world software work, Codex is positioning itself not as a novelty but as infrastructure.

If you’re a developer, the practical takeaway is simple: this isn’t just about writing code faster anymore. It’s about whether you want an always-on companion that can reason about your code, operate your stack, keep track of your projects, and quietly take care of the work you didn’t really want to do in the first place. OpenAI’s bet is that once you get used to that, going back to a pre-Codex workflow will feel as strange as going back to a text editor without search. And judging by the early usage and reviews, a lot of teams are already there.