What Codex is
Codex is OpenAI's AI coding product — an agent that can write, fix, and ship software on your behalf. Think of it less like a smart autocomplete and more like a junior developer you can hand a task to and walk away from: it works through the problem, uses tools, browses the web, operates your computer if needed, and comes back with results.
It lives in several places at once: a standalone app, a command-line tool (CLI), extensions for popular code editors, and — as of mid-2026 — a unified tab inside the ChatGPT desktop app for macOS and Windows.
A quick history
The name "Codex" goes back to 2021, when OpenAI published research introducing a model trained specifically on code, along with a benchmark called HumanEval for measuring how well AI could write working programs. That early model became the engine behind GitHub Copilot's first version.
The product you use today is a different beast. OpenAI relaunched Codex in May 2025 as a full developer platform, not just a model. Since then it has received a rapid series of upgrades — GPT-5-Codex, GPT-5.1-Codex-Max, GPT-5.3-Codex, and now GPT-5.6 — each generation faster and more capable of handling complex, multi-step work with less hand-holding.
Why it matters (the "so what")
Most coding tools help you write one function or fix one bug at a time. Codex is designed for longer work: tasks that take minutes or hours, span multiple files, require searching the web for documentation, or involve running and testing code in a real environment.
That shift matters because a lot of real software work isn't a single question — it's a project. Codex can:
- Write and refactor code across an entire codebase
- Review pull requests and flag security issues
- Operate your computer — clicking, typing, and navigating apps — to reproduce bugs or test a running app visually
- Build and deploy web apps directly from the Codex app, hosted by OpenAI
- Run as a background agent in the cloud, continuing work while your laptop is locked
How it's built
Codex runs on a dedicated family of models — not the same GPT you use for chat. These Codex-specific models are tuned to handle long coding sessions, adjust how hard they "think" based on task complexity (quick for simple questions, sustained for hard problems), and work well inside agentic loops where they call tools, check results, and try again.
A speed-optimized variant called GPT-5.3-Codex-Spark, built in partnership with chip company Cerebras, can generate over 1,000 tokens per second — fast enough to feel like real-time collaboration rather than waiting for a response.
Who's using it and how
OpenAI reported 4 million weekly active users by April 2026. Internally, OpenAI's own teams saw Codex output grow 56× in Research and 27× in Engineering between November 2025 and mid-2026 — a sign of how quickly even the people building these tools are leaning on them.
On the enterprise side, Samsung Electronics deployed Codex to employees worldwide. Consulting giants Accenture, PwC, and Infosys signed on as partners through a program called Codex Labs. Codex is also available on Amazon Web Services and inside Cloudflare's Agent Cloud, so companies can run it within their existing infrastructure rather than sending data to a separate service.
Getting into the details: what's new
A few recent additions are worth knowing about:
- Computer Use: Codex can now see and control Windows and Mac desktops, letting it interact with apps the way a human would — useful for testing, visual bug reproduction, or automating workflows that don't have an API.
- Sites: A built-in feature for building and deploying lightweight web apps and internal tools, hosted by OpenAI, without leaving the Codex app.
- Secure MCP Tunnel: An enterprise feature that lets Codex connect to private internal servers without exposing them to the internet — a key requirement for regulated industries.
- Goal mode: A stable "just get it done" mode where you describe an outcome and Codex figures out the steps, rather than waiting for you to approve each one.
OpenAI also acquired Astral — the company behind the Ruff Python linter and the uv package manager, two widely-used Python developer tools — specifically to deepen Codex's Python ecosystem.
The bigger picture
By July 2026, Codex was being described as a ChatGPT "superapp" — the coding surface of a unified client that also handles chat and productivity work. That framing signals where OpenAI is heading: not separate tools for separate tasks, but one agent that handles your whole workday, with Codex as the software-development arm.
Its main rival in this space is Claude Code from Anthropic, which competes on similar agentic coding capabilities. Cursor's Composer is another specialist option, notable for being significantly cheaper per task. The competition is pushing all of these products to improve quickly.
