OpenAI Codex CLI Review 2026: Features, Pricing & Verdict

I have been using OpenAI's Codex CLI as my daily driver for almost four weeks now. Long enough to ship two production features with it, blow up one staging branch, and form an opinion that does not consist of the phrase "it depends."

This is the review I wish I had when I first heard about Codex CLI. No paid placements, no breathless launch hype, no recycled press release bullet points. Just what it does, what it costs, where it beats Claude Code, where it loses, and whether I think you should install it tonight.

Short answer: yes. Slightly longer answer below.

What is Codex CLI?

Codex CLI is OpenAI's lightweight terminal coding agent. You install it with one npm command, sign in with your ChatGPT account, and run codex inside any folder. From there you describe what you want in plain English and the agent reads your code, edits files, runs shell commands, runs your tests, iterates if they fail, and finally hands you a diff to approve.

It is open source, lives at github.com/openai/codex under the Apache 2.0 license, and as of April 2026 it has crossed 41,000 GitHub stars. The repo is the actual production code, not a thin wrapper, and you can read every prompt the agent sends if you want to.

The brain inside is a model called codex-1. OpenAI describes it as a version of o3 trained with reinforcement learning on real-world software engineering tasks. In practice that means it generates code that mirrors human style and pull-request preferences, follows instructions carefully, and is willing to keep iterating until tests pass instead of stopping at the first plausible answer.

There is a cloud-hosted "Codex Web" version too, but this review is about the terminal CLI specifically — that is the part that competes head-to-head with Claude Code, Cursor's agent mode, and Aider.

Six features that actually matter

Every coding agent ships a feature checklist a mile long. Here are the six things I noticed Codex CLI does differently after using it daily.

The combination of the sandbox plus codex-1 plus the auto-PR loop is what differentiates Codex CLI from Claude Code in practice. Claude Code is happier as a pair-programmer that you watch type. Codex CLI is happier as a contractor you give a ticket to and check on later.

How to install and use Codex CLI

This is the part where most reviews get unnecessarily long. The actual install is two commands.

Mac users can swap step one for brew install --cask codex if you prefer Homebrew. Windows works through WSL or, as of the March 2026 update, natively through PowerShell with the same npm command.

Once you are inside an interactive session, the prompts that get the best results in my experience look like this:

Three things that command does well: it tells the agent which file to read first, it gives it a verification step (the existing tests), and it ends with the PR title so the agent does not have to guess. Codex follows that structure cleanly. I have wasted hours trying to get other agents to do the same thing in three iterations.

In-session slash commands worth knowing on day one:

• /model — switch between codex-1, GPT-5, GPT-5-Codex, and codex-mini.
• /status — see how much of your daily quota you have left.
• /sandbox — toggle the cloud microVM on or off.
• /approval suggest|auto-edit|full-auto — change how aggressive the agent is.
• /mcp — list and configure attached MCP servers.

Pricing: what you really pay

Codex CLI's pricing story is simple in a way most AI coding tools refuse to be. There is no separate Codex bill. You pay for ChatGPT, you get Codex.

There is also an API path through codex-mini-latest at $1.50 per million input tokens and $6 per million output tokens, with a 75% prompt-caching discount on cached reads. That is the route to take if you want to embed Codex in your own scripts or CI pipeline rather than driving it from the terminal.

The honest pricing comparison: on Plus I can run Codex hard for an entire workday and rarely hit a limit. On Claude's $20/month Pro tier I burn through my limit on a handful of complex prompts before lunch. That four-times-fewer-tokens efficiency claim from OpenAI lines up with my own usage logs.

Pros and cons after a month

The cons list is honest, not theatrical. None of them stopped me from making Codex CLI my default. They did stop me from uninstalling Claude Code.

Alternatives: Codex CLI vs Claude Code, Cursor, Copilot, Windsurf

If you are reading this you are almost certainly comparing Codex CLI against one of these four. Here is the honest grid.

If you want a deeper teardown of just the Codex vs Claude question, I also wrote a longer head-to-head over at Codex CLI vs Claude Code 2026, including benchmark numbers and side-by-side prompt outputs.

A separate datapoint worth knowing: a 500+ developer Reddit survey published in early 2026 found that 65.3% of respondents said they had switched their primary daily-driver agent to Codex CLI, with weighted upvotes favoring it at 79.9%. That is the kind of community shift that does not happen unless something is meaningfully different.

Frequently asked questions

Final verdict: 4.3 / 5

Codex CLI is the first OpenAI coding tool in years that I have not had to apologize for recommending. The codex-1 model is genuinely tuned for finishing tasks instead of producing plausible-looking starter code. The cloud sandbox is the right default for autonomous runs. The pricing structure — bundled free with the ChatGPT Plus I was already paying for — eliminates the awkward second subscription decision.

It is not perfect. Claude Code still wins for front-end work and complex architectural reasoning, and the Claude MCP and skills ecosystems are bigger today. But for the 70% of my week that consists of "read this code, change this thing, run the tests, open a PR," Codex CLI is now the tool I open first.

If you are paying for ChatGPT Plus and you have not run npm install -g @openai/codex yet, you are leaving free productivity on the table. That is the whole review.