The latest Codex release isn’t just a developer tool update — it’s a stealth preview of ChatGPT’s upcoming super app, and it changes what autonomous AI co-workers can actually do.
On April 16, 2026, OpenAI rolled out a major update to Codex — and most people either missed it entirely or dismissed it as just another incremental feature bump for developers. That would be a significant mistake. What landed last week is arguably one of the largest changes in how AI works that we’ve seen in the past three years. This isn’t a new model release or a chatbot upgrade. It’s a fundamental shift in how AI agents operate on your computer — the first true autonomous co-worker that can actually work alongside you, not instead of you.
To understand the magnitude of what’s happening, you have to zoom out. ChatGPT is going through a massive transformation. Consumer and enterprise momentum is moving rapidly toward autonomous desktop operations — AI that doesn’t just answer questions but actually does work on your machine. OpenAI has been building a super app in plain sight, and they’ve now confirmed it. In a recent Engadget interview, OpenAI’s head of Codex, Tibo, stated bluntly: “We’re building the super app out in the open.” That statement alone should make everyone pay attention.
Until recently, OpenAI maintained three separate desktop experiences: the ChatGPT app for general conversation and reasoning, the Atlas agentic browser for web-based tasks, and Codex for developer workflows. Now they’re merging everything into one unified platform. The walls between these products are coming down. And the barrier to entry? Any paid ChatGPT plan. That’s it. What was once fragmented across three apps is converging into a single workspace — and the April 16 update is the clearest signal yet of where this is all heading.
The Critical Difference: Operating System-Level Autonomy
This is the single most important thing to understand about the updated Codex, and it’s the detail that separates it from virtually every competitor on the market: OpenAI’s Codex operates at the operating system level, not the user interface level. That distinction sounds technical. It is anything but. It changes everything about how practically useful an AI agent can be in your daily work.
Most computer-using agents today — including Anthropic’s Claude Code and Claude Co-work — work by essentially screen-jacking your computer. They take over your mouse, your screen, your browser tabs. They click where they need to click, type where they need to type, and navigate where they need to navigate. The problem? You can’t work while they work. Your computer becomes theirs. You sit and watch, or you walk away and hope nothing goes wrong. They’re impressive technology demos, but they’re impractical for real co-working during business hours.
OpenAI’s updated Codex changes this entirely. It operates natively within macOS (and Windows) at the OS level. It has its own independent cursor. You can literally see it working alongside you — navigating apps, clicking, typing — while you continue doing your own work in your own applications. Multiple agents can run simultaneously without interrupting your workflow. It doesn’t screenjack. It doesn’t freeze your browser. It doesn’t commandeer your mouse. You keep working. It keeps working. Independently. In parallel.
Why This Matters for Real Productivity
With Claude’s computer use, the only realistic time to run complex agent workflows was overnight — and even then, a single error or permission prompt could derail everything while you slept. With Codex, if something comes up, you can address it immediately because you’re sharing the workspace. This is the first true unlock of consistent AI productivity for everyday users.
What’s New in This Codex Update
The April 16 update expanded Codex in several directions simultaneously. Understanding these individual features is important, but what matters even more is understanding how they combine — because the real power isn’t in any single capability, it’s in the compound effect of all of them working together inside one application.
Background Parallelism
The headline feature is background computer use at the OS level. As described above, Codex now has its own independent cursor on Mac, allowing it to operate apps, click buttons, fill forms, and navigate windows while you do your own work uninterrupted. But it goes further: multiple agents can run simultaneously. You can have one agent researching in a browser, another editing a document, and a third testing a frontend design — all in parallel, all in the background, all without touching your own workflow. This is not a gimmick. This is a paradigm shift in what “using AI” means on a desktop.
Integrated Browser and File Viewer
Codex now includes a built-in web browser and file browser that live directly inside the application. The in-app browser lets you comment directly on web pages to provide precise instructions to the agent — enormously useful for frontend development, design iteration, and debugging. The file viewer provides rich previews for PDFs, spreadsheets, slides, and documents, along with a summary pane that tracks agent plans, sources, and artifacts. The net effect is a tighter feedback loop: instead of constantly switching between Codex and other apps to check results, everything lives in one workspace.
Images 2 Model Access
Codex can now generate and iterate on images using OpenAI’s latest gpt-image-1.5 model directly inside the app. Combined with screenshots and code, this makes it possible to create visuals for product concepts, frontend designs, mockups, and even game assets inside the same workflow where you’re building the actual product. For designers and product managers, this collapses what used to be a multi-tool, multi-step process into a single continuous workflow.
90+ New Plugins
OpenAI released more than 90 additional plugins with this update, including integrations with Atlas, GitLab, Gmail, Slack, Atlassian Rovo (for JIRA management), CircleCI, CodeRabbit, Microsoft Suite, Neon by Databricks, Render, and many more. Critically, these aren’t just thin API wrappers. OpenAI frames them as packaging units that can combine skills, app integrations, and MCP servers (Model Context Protocols). Custom MCPs are also supported, meaning teams can build their own integrations tailored to internal tools and workflows.
Chronicle (Experimental)
Released just days after the main update on April 20, Chronicle is perhaps the most ambitious — and most controversial — feature in the new Codex. It’s essentially a screen memory feature that augments Codex’s memory with context from your screen activity. If you’ve been following the AI industry, this will sound familiar: it’s functionally similar to Microsoft’s controversial Windows Recall feature, but implemented more quietly inside Codex as a research preview.
Here’s how it works: Chronicle runs sandboxed background agents on your Mac that build memories from screen captures. This enables prompts like “finish what I was doing” or “update that dashboard from earlier” — Codex actually knows what “this” and “that” refer to because it saw them on your screen. Over time, it learns which tools you use, which projects you return to, and which workflows you rely on. Currently available only to Pro subscribers ($200/month) on Mac, Chronicle requires granting macOS Screen Recording and Accessibility permissions. The privacy implications are significant: screen captures are stored temporarily on-device (deleted after six hours), but the generated memories are stored as unencrypted markdown files locally. OpenAI warns that Chronicle also burns through rate limits quickly.
Persistent Memory and Proactive Automations
Memory now persists across sessions in Codex, remembering your preferences, tech stack, project conventions, and recurring workflows. But the real unlock is what happens when you combine persistent memory with Chronicle and the new proactive automations feature. Codex now notices when you perform the same manual tasks repeatedly and proactively suggests automations — it will essentially say, “I’ve noticed you do this every morning. Want me to handle it automatically?” The combination of persistent memory + Chronicle + suggested automations has been described as one of the stickiest feature combos for power users. Once Codex knows how you work, it starts working for you without being asked.
Availability
The updated Codex is available now on Mac and Windows (some features are Mac-only for now, with Windows parity pending). The barrier to entry is low: any paid ChatGPT plan grants access. This is notable because Anthropic has reportedly been considering removing Claude Code from their base plan, while OpenAI leadership has committed to keeping Codex available on all paid tiers. The platform has already crossed 4 million users — a remarkable adoption curve for a tool that launched its desktop app less than three months ago.
Five Steps to Get the Most Out of Codex
Step 1: Connect Your Tools First
Before you write a single prompt, spend fifteen minutes connecting every tool you use daily. Whether it’s Slack, Gmail, Notion, JIRA, or any of the 90+ available plugins — get everything connected from day one. Here’s the key insight many people miss: when combined with background computer use, any desktop application essentially functions like a connected app, because Codex can navigate it autonomously. But native plugin connections are faster, more reliable, and more secure. The more context Codex has from the start, the smarter its suggestions become immediately. This also unlocks scheduling capabilities — you can’t automate a workflow that touches Gmail if Gmail isn’t connected.
Step 2: Port Your Top Prompts and Run Them Side by Side
Don’t take anyone’s word for how good (or bad) Codex is — run your own use cases. Take your three best-performing workflows from competing tools (Claude Co-work, Claude Code, Cursor, or whatever you currently use) and three workflows that haven’t worked well elsewhere, and test them head-to-head in Codex. Benchmarks published by AI companies are useful for general guidance, but they tell you nothing about your specific tasks, your specific data, and your specific workflow patterns. Real-world results on your work are what matter. You may be surprised — some tasks that failed in other tools may succeed here because of the OS-level approach, and vice versa.
Step 3: Audit Every Workflow That Involves Clicking or Browsing
Most people dramatically underestimate how much time they spend on mundane clicking, dragging, tab-switching, and navigating tasks. Here’s a powerful hack for identifying automation opportunities: record a quick screen video of yourself performing a repetitive workflow, narrating what you’re doing as you go. Then upload that video to an AI tool with video understanding (such as Google’s AI Studio using a Gemini Flash model) and have it generate a step-by-step tutorial based on both the transcript and screen activity. Feed that tutorial directly to Codex as a prompt. Prioritize workflows that are low-risk (mistakes won’t cause real damage), high-frequency (you do them daily or weekly), and highly scalable (the time savings multiply over weeks and months).
Step 4: Let Memory Replace Your Setup Time
Stop re-explaining your preferences every session. This is one of the most common productivity drains with AI tools — the context reset. Every new conversation starts from zero, and you spend the first five minutes telling the AI who you are, what you’re working on, and how you like things done. With Codex’s persistent memory, combined with Chronicle and suggested automations, your setup time approaches zero over time. Codex learns your tech stack, your coding conventions, your design preferences, and your communication style. It carries all of that forward automatically. Lean into this: explicitly tell Codex your preferences early, and let the system compound that knowledge across sessions.
Step 5: Schedule One Recurring Task After Iterating
The goal of all the above steps is to get here: a consistent, verifiable automated workflow that runs on a schedule. But don’t rush it. First, always study the chain of thought to understand how Codex is processing your requests. Run the same prompts multiple times — generative AI produces different results each time, and you need to understand the variance. Look at as much reasoning as possible, identify where things go wrong, refine your prompts, and only then set up a recurring schedule. The payoff is enormous: imagine waking up to find that a complex multi-step task — pulling data, cross-referencing sources, generating a report, formatting it, and sending it to stakeholders — was completed overnight, every night, without your involvement.
Real-World Demo Insights
To stress-test the new Codex, consider a complex multi-app workflow designed to push every new capability simultaneously. The test: have Codex navigate to Twitter/X via computer use to find trending AI stories, cross-reference those stories against a daily industry newsletter via the connected Gmail plugin, select the most interesting topic, then open the Claude desktop app via computer use (not a plugin — Codex’s computer use actually detected Claude as an installed desktop application and made it accessible), instruct Claude to search for additional information on the topic, and finally use a connected Canva plugin to create a three-page visual explainer.
The results are revealing. The speed of computer use in Codex is dramatically faster than Claude’s computer use — tasks that take Claude 10+ minutes of painstaking screen navigation can be completed by Codex in a fraction of that time. However, the system isn’t perfect. In testing, the agent sometimes used the wrong version of an app — for example, opening gmail.com in the in-app browser instead of using the connected Gmail plugin — suggesting that prompts need to be extremely specific and tested iteratively. Using the GPT-5.3 Codex Spark model offers faster execution speed but noticeably less accuracy compared to higher-tier models. The Gmail API connection was sometimes temperamental, requiring retry logic.
But the bigger lesson transcends any individual test result: the goal of running these complex multi-app workflows isn’t just to complete them once. It’s to iterate on the prompts, study the chain of thought, identify failure points, and refine until the workflow is reliable enough to schedule as a recurring automation. That’s the endgame. By investing time upfront in testing and prompt refinement, users can build a genuine autonomous workflow pipeline — a set of tasks that run daily or weekly, producing consistent, verified outputs without manual intervention.
The Competitive Landscape
Anthropic jumped to an early lead in the desktop AI agent category with Claude Code, and later expanded into Claude Co-work — a product born from observing that people were using Claude Code for non-technical tasks far beyond its original scope. Microsoft has Copilot Co-work embedded across its Office ecosystem. Google has upcoming updated agent capabilities. Open-source tools like OpenClaw have become some of the most popular open-source software ever. The market is crowded, the competition is fierce, and everyone is racing toward the same destination: an AI that does real work on your real computer.
But OpenAI’s approach with Codex — particularly the OS-level autonomy rather than screen-jacking — may leapfrog all of them for practical, daily use. There’s a deep irony here: Anthropic may actually be behind in the desktop race despite being first, because their approach requires exclusive access to your computer while the agent works. OpenAI’s Codex lets you share your workspace. For non-technical business leaders, this is a game-changer. The name “Codex” is arguably the product’s biggest liability — it suggests a developer-only tool, when the reality is that everyday business users can benefit enormously from this technology right now. Whether OpenAI rebrands it or merges it fully into ChatGPT, the underlying capability is platform-defining.
Looking Ahead: The Super App Is Already Here
The super app is coming — or more accurately, it’s already arriving in pieces. Whether OpenAI keeps the Codex name or folds everything into a unified ChatGPT experience, the foundation is here and usable today. OS-level autonomy. Persistent memory. Proactive automations. Screen-aware context via Chronicle. Ninety-plus plugin integrations. Background parallelism. An in-app browser. Image generation. For anyone who has spent the past three years waiting for AI to become a true co-worker rather than just a sophisticated chatbot, this is the moment to start experimenting seriously.
The combination of these capabilities doesn’t just represent a feature update — it represents a fundamentally new way of working with AI. The gap between “AI that answers questions” and “AI that does work” has been closing for years. With this Codex update, for the first time, it feels like that gap has meaningfully closed. The agents aren’t perfect. The prompts require iteration. The privacy trade-offs are real. But the trajectory is unmistakable. The desktop AI co-worker isn’t a future product announcement. It’s running right now, with its own cursor, on millions of machines — and it’s getting better every week.
Leave a comment