Codex is OpenAI’s agentic coding product — the thing on the other side of the Claude Code / Cursor / Codex triangle that most serious AI coding teams now use together rather than picking one. For a long stretch of 2025, Codex was a CLI tool aimed at developers. That’s not what it is anymore. As of April 2026, Codex is a desktop app, a browser-driving agent, a schedule-your-automations tool, and a coding assistant — all bundled into the ChatGPT subscription you probably already pay for.

New in April 2026: GPT-5.5 and browser use

On April 23, 2026, OpenAI made GPT-5.5 the default model inside Codex. That matters because GPT-5.5 is the first fully retrained base model since GPT-4.5, and it hit 82.7% on Terminal-Bench 2.0 — state-of-the-art on agentic, multi-step workflows. In practical terms: Codex is now noticeably better at finishing tasks it used to abandon halfway through, and noticeably better at recovering from its own mistakes without asking you to restart.

The other April 2026 change is browser use. Codex can now operate a browser inside the desktop app — click through pages, log into apps, run through checkout flows, capture screenshots, iterate until a task is done. That’s on top of the April 16 “Codex for (almost) everything” update that added macOS computer use, 90+ plugins (Jira, Notion, Slack, the Microsoft 365 suite), image generation, persistent memory, and scheduled automations. Read the GPT-5.5 launch guide for what’s in the release and what it means for vibe coders.

OpenAI cited 3 million weekly active Codex users in April 2026, up from 1.6 million in early March. A chunk of that growth is non-developers — marketers, ops people, finance analysts — using Codex the way Claude Code gets used by developers, but on business workflows rather than codebases.

What Codex actually is

Codex is a desktop app (Mac and Windows) that you log into with your ChatGPT account. Inside the app, you describe what you want in plain English and Codex figures out how to do it. If the task involves writing code, it writes code. If it involves opening your browser and doing something, it drives the browser. If it involves pulling data from Slack and writing a summary in Notion, it calls those plugins. The interface is intentionally close to ChatGPT — you type, it responds, there’s a running conversation.

The CLI version still exists (and is free and open-source). Most people don’t use it. The desktop app is what the 3M weekly active users number refers to.

The pitch to non-technical founders is: instead of stitching together Zapier, a no-code builder, a code editor, and four other tools, run the whole thing through one agent. That pitch is half true. It works well for ad-hoc multi-step tasks — “pull my last 30 days of Stripe revenue and draft an investor update” — and less well for repeatable production workflows.

Where Codex is actually good

The coding is solid. With GPT-5.5 as the engine, Codex can take a vague description of a bug and often fix it end-to-end, running the tests, looking at errors, and iterating. It’s particularly strong at methodical debugging — if you show it a failing test, it will systematically try to narrow down the cause rather than guess. Developers comparing it against Claude Code generally say: Claude Code is faster at raw codegen, Codex is better at figuring out what’s actually wrong.

The browser use is a genuine differentiator. If you want an agent that can log into your live app and verify that the signup flow works, Codex is the one that can do that today. Cursor can’t. Lovable can’t. Claude Code technically can via MCP plugins, but not as smoothly.

The 90+ plugin ecosystem makes Codex surprisingly competent at the “business operations” tier of work. Drafting Jira tickets from a Slack thread. Pulling a Notion doc and turning it into a deck outline. Cross-referencing data across apps without you writing any integration code. For founders running a small company themselves, this is probably a better use case than the coding features.

Where Codex is not the right tool

It’s not a visual app builder. If you don’t want to see any code and you want to describe an app and get a working product back, use Lovable, Bolt, or Base44. Codex will generate the code, but you’ll end up in a file tree you can’t read.

It’s not an IDE. If you’re actively writing code and want inline AI assistance, Cursor or Windsurf are purpose-built for that experience and Codex is not. Codex’s sweet spot is “I described a task and the agent went off and did it” — not “I’m writing code and need real-time help.”

It’s not free. The CLI is free; the desktop app requires a paid ChatGPT plan. The usage limits on Plus are real. If you’re running browser automation or computer-use tasks regularly, you’ll push those limits within a week.

Pricing

Codex is included with any paid ChatGPT plan: Plus ($20/mo), Go, Team, Pro ($200/mo), Business, and Enterprise. Plus is the entry point for individual use. Pro gives you 20x the Plus usage limits — worth it if Codex becomes your daily driver, not worth it if you’re just exploring. There’s no standalone Codex subscription; you’re paying for ChatGPT and Codex comes with it.

Through May 31, 2026, paid plans get 2x Codex usage as a rollout bonus. Plus plans get 2x standard limits; Pro gets 10x instead of the normal 5x.

Who this is actually for

If you already pay for ChatGPT Plus or Pro, you effectively already have Codex. The question is whether it’s worth the hour of setup and the week of trying to get it to do what you want.

For a non-technical founder, Codex is worth the afternoon if: you’re comfortable watching an agent do things on your behalf, you have a handful of recurring multi-step workflows that span your tools, and you’d rather pay $20/mo for one agent than $200/mo across four SaaS tools. If those don’t describe you, stick with the tool you’re using.

For a technical founder or a PM with coding experience, Codex is worth taking seriously alongside Claude Code. The practical pattern that’s emerging: Claude Code for deep codebase work, Codex for browser-driven and cross-app tasks. Comparing them head-to-head is a losing framing — they’re good at different things.

Bottom line

Codex is no longer the developer CLI it was six months ago. The April 2026 updates — computer use, browser use, plugins, scheduled automations, and now GPT-5.5 — make it a legitimate “do my work” agent rather than a coding assistant. The pricing is fair if you’re already on ChatGPT. The non-technical founder fit is real but not universal — it assumes you’re comfortable supervising an agent rather than clicking through a UI. If that’s you, Codex is worth the afternoon. If it isn’t, there are better starting points for your first vibe coding project.

Was this helpful?

OpenAI Codex