Most coding assistants live inside your editor. They sit beside your cursor, suggest the next line, and wait for you to accept or reject. OpenAI Codex does something different: it takes the work off your machine entirely. You describe what needs doing, and a cloud-based agent writes, runs, and tests the code in its own sandboxed environment. When it finishes, you review a pull request. The shift is not just ergonomic — it changes who controls the execution loop.

Codex is not a single product. It is two products that share a name and complement each other. Understanding the split is important because the trade-offs are different for each.

What Is OpenAI Codex

Codex is OpenAI’s dedicated coding agent, available both as a cloud application inside ChatGPT and as a standalone open-source command-line tool. The cloud version launched in mid-2025 and has evolved quickly since then. The CLI followed shortly after under the Apache 2.0 licence, meaning anyone can inspect, modify, and redistribute its source code.

The core idea behind both is the same: rather than autocompleting lines of code while you type, Codex operates as an autonomous agent. You give it a task — fix this bug, add this feature, refactor this module — and it plans, writes code, creates or modifies files, runs tests, and reports back. It does not need you to hold its hand through each step, though you can intervene whenever you want.

What makes Codex distinct from tools like Copilot or Cursor is that it was designed from the start to work outside your editor. The cloud version runs in its own isolated environment with its own file system. The CLI version runs in your terminal. Neither requires you to have an IDE open. For developers who spend most of their time in the terminal, and for teams who want to dispatch coding tasks the way they dispatch CI jobs, this is a meaningful architectural choice.

The agent can browse documentation, install dependencies, and execute shell commands inside its sandbox. When it finishes, the result is a diff you can review, not a block of text you need to copy and paste somewhere.

The App And The CLI

The two versions of Codex serve different needs and work differently in practice.

The Codex App is the cloud-hosted version, accessible through the ChatGPT interface. It connects to your repository — GitHub is the primary integration — and spins up isolated worktrees for each task. A worktree is a separate working copy of your codebase, which means the agent can modify files without affecting your local branch or anyone else’s work. You can run multiple agents in parallel, each working on a different task in its own worktree. One agent refactors a module while another writes tests for a different part of the system. The results come back as pull requests you merge when ready.

The app also includes an in-app browser, which lets the agent read documentation, check API references, or verify that a web application renders correctly after making changes. This is not a gimmick — coding agents that cannot see what a page looks like after modifying its markup are guessing at the result. The browser closes that gap.

Since March 2026, the Codex App has included full Windows support, which was a notable gap in the first year. Scheduled long-running tasks are also available: you can set a job to run overnight, processing a large refactor or a test suite that takes hours, and review the results in the morning.

The Codex CLI is the open-source counterpart. It runs in your terminal, authenticates using your ChatGPT account by default (no separate API key needed, no extra charge beyond your existing subscription), and works against your local file system. If you prefer using API keys directly — for billing separation or to access specific models — that option is available too.

The CLI is intentionally minimal. It does not manage worktrees or orchestrate parallel agents. What it does is give you a fast, scriptable interface to the same underlying models. You type a task, the agent works through it in your terminal, and you see the changes in real time. For developers who live in the command line, this is the more natural fit. And because it is open source under Apache 2.0, you can extend it, wrap it in your own tooling, or audit exactly what it does.

Skills And Extensibility

Where Codex becomes more than a coding assistant is in its skills system. Skills are reusable instruction packages that tell the agent how to handle specific types of tasks. Think of them as specialised playbooks: a skill might define how to set up a new microservice in your stack, how to write tests according to your team’s conventions, or how to handle database migrations with your ORM of choice.

You can write your own skills, share them across a team, or use community-contributed ones. This turns Codex from a general-purpose coding agent into something that understands the specific patterns and standards of your project. A new team member with the right skills loaded can dispatch tasks that follow the team’s existing conventions without having memorised them yet.

Memory is another feature that changes how the agent behaves over repeated sessions. Currently in preview, memory allows Codex to retain context about your project across sessions — architectural decisions, naming conventions, common pitfalls, preferred libraries. Instead of re-explaining your stack every time, the agent accumulates understanding over time. This is still early, and the practical limits of how much context it retains are being tested, but the direction is clear: a coding agent that knows your project is far more useful than one that starts from scratch every morning.

The browser integration in the cloud app extends the agent’s capabilities beyond code. It can verify that a deployed endpoint returns the expected response, check that a UI change renders properly, or pull information from internal wikis and documentation sites. Combined with skills and memory, this creates an agent that operates across the full surface of a development workflow, not just the text editor.

Pricing And Models

Codex offers three model tiers, each suited to different workloads. Pricing is per million tokens (input/output):

GPT-5 Codex at $1.25 / $10 is the workhorse. It handles most coding tasks competently — feature implementation, bug fixes, test writing, refactoring. For the majority of day-to-day development work, this is the model to use.

GPT-5.1 mini at $0.25 / $2 is the budget option. It is faster and cheaper, well-suited for simpler tasks like formatting, boilerplate generation, straightforward code translations, or high-volume batch jobs where cost matters more than nuance.

GPT-5.3 at $1.75 / $14 is the most capable model, designed for complex reasoning tasks — intricate debugging, architectural planning, working through problems that require holding many constraints in mind simultaneously. The cost is higher, and for routine work the difference may not justify the price, but for hard problems it is noticeably better.

The CLI uses your ChatGPT subscription by default, which means Pro and Plus subscribers can use it without additional per-token charges. This makes the CLI effectively free for anyone already paying for ChatGPT — a genuine advantage over competitors that charge separately for API access and chat subscriptions.

The coding agent landscape in 2026 is crowded. Claude Code, Gemini CLI, GitHub Copilot Workspace, Cursor — all capable, all improving rapidly. What Codex brings to the table is a specific combination: a cloud-native architecture that runs agents in parallel without touching your machine, an open-source CLI that respects developers who prefer transparency and control, and a skills system that lets you encode your team’s knowledge into reusable instructions. Whether that combination fits your workflow depends on how you work. But the fact that the CLI is open source and authenticates through an existing ChatGPT subscription means the cost of trying it is essentially zero. That is a strong starting position.