OpenAI Codex: What It Is and How to Use It
In short
- OpenAI Codex is a family of AI coding agents from OpenAI: a cloud agent inside ChatGPT, the open-source Codex CLI for the terminal, and extensions for VS Code and JetBrains. Everything runs off a single ChatGPT account.
- The cloud Codex clones your GitHub repository into an isolated container on its own, writes code, runs tests and opens a pull request — several tasks in parallel.
- Access comes through a ChatGPT subscription; Codex isn't sold separately. As of July 2026: Plus is $20/month, Pro is $100–200/month, Business starts at $20 per seat; on the free tier the limits are token gestures.
- As of July 2026 Codex runs on the GPT-5.6 model family (Sol, Terra, Luna), which OpenAI began rolling out on July 9, 2026.
- Strengths: parallel work in the cloud, deep GitHub integration, a low price of entry. Weaknesses: Plus limits burn out fast on big tasks, and from Russia both access and payment come with caveats.
- Codex reads AGENTS.md — an open instruction format for agents. Keep your project rules separate from the tool: the qvib engine exports to Claude Code, Cursor and AGENTS.md, so you can try Codex without losing your accumulated work.
What OpenAI Codex is
OpenAI Codex is an agentic development environment: a set of tools where the AI doesn't just suggest lines of code but performs tasks on its own — reads the repository, edits files, runs commands and tests, and packages up the changes. It's a textbook example of the agentic approach we covered in detail in what agentic coding is.
OpenAI has used the name Codex for two different products. The first was the 2021 model that powered early GitHub Copilot. The second, current one was relaunched in 2025 as a full agent and has since grown into an umbrella product: a cloud agent in ChatGPT, a console CLI, IDE extensions, a GitHub bot, and desktop apps for macOS and Windows. All surfaces use shared models and the shared context of your account.
As of July 2026 Codex runs on the GPT-5.6 family — Sol (flagship), Terra (balanced) and Luna (fast and cheap). The family started rolling out on July 9, 2026; on paid plans you pick the model and the effort level manually.
How Codex is built: CLI, IDE and the cloud agent
Codex has several "surfaces", and they cover different scenarios.
Codex CLI is an open-source agent for the terminal: written in Rust, installed with npm install -g @openai/codex. It works inside your local project folder: reads code, edits files, runs tests, makes commits. Security rests on two layers: the sandbox (what the agent can technically do — where it can write, whether it has network access) and the approval policy (when it must ask you). Modes range from "propose every action" to fully autonomous.
The cloud agent is Codex inside ChatGPT: you describe a task in text, the service clones your connected GitHub repository into an isolated container on OpenAI's infrastructure, writes code, runs tests and opens a pull request. Tasks run in parallel and don't tie up your computer.
AGENTS.md is a file with project instructions for the agent: stack, rules, build commands, prohibitions. It's the counterpart of CLAUDE.md in Claude Code, but in an open format that other tools understand too — that's what portability is built on, more below.
| Surface | Where it runs | What it can do | When to choose it |
|---|---|---|---|
| Codex CLI | Your terminal | File edits, tests, commits, MCP | Local work, full control |
| IDE extension | VS Code, JetBrains | The same plus edits right in the editor | Your usual IDE flow |
| Cloud agent | ChatGPT (web and desktop) | Parallel tasks, PRs on GitHub | Background and long-running tasks |
| GitHub bot | Comments in PRs | Code review, tasks via @codex |
Team code review |
How is Codex different from ChatGPT?
The short answer: ChatGPT is a conversationalist, Codex is a doer. In regular chat you get advice and snippets and carry them into the project by hand. Codex works with the repository directly: it sees the project structure, changes several files at once, runs linters and tests, fixes a broken build itself, and packages the result as a diff or a pull request.
At the same time it's one product: a ChatGPT subscription gives you both chat and Codex, and limits are counted from a shared credit pool. There's no separate "Codex subscription" to buy.
How to get access to Codex
- Create a ChatGPT account. Codex is formally available on every plan, including the free one, but with minimal limits — for real work you need Plus or above.
- Cloud agent: open the Codex section on chatgpt.com, connect GitHub and grant access to the repositories you need.
- CLI: install the
@openai/codexpackage via npm (Node.js 22+ required) and sign in with your ChatGPT account — or with an API key, in which case you pay per token through the OpenAI platform. - IDE: install the VS Code or JetBrains extension and sign in with the same account.
On Russia: OpenAI services are officially unavailable there, so the legal routes are few — a foreign bank card issued in your name, or payment through intermediary services, with the risks understood: markups, no guarantees, and sometimes handing account access to third parties. For API work there are aggregators that accept Russian cards. We covered the same problem for Claude in Claude API from Russia — the logic is identical.
How much Codex costs
On April 2, 2026 OpenAI moved Codex to a token model: plan limits are counted in credits tied to actual token consumption rather than to a number of messages. Prices as of July 2026:
| Plan | Price (as of July 2026) | What it gives you for Codex |
|---|---|---|
| Free | $0 | Token-gesture limits, entry-level models |
| Go | $8/month | A small credit pool |
| Plus | $20/month | Base pool, choice of GPT-5.6 models |
| Pro | $100 or $200/month | ×5 or ×20 the Plus limits |
| Business | from $20 per seat/month | Team pools, admin controls |
| Enterprise | On request | Custom limits and security |
If you sign in to the CLI with an API key, usage is billed at platform rates (as of July 2026): GPT-5.6 Luna — $1/$6 per million input/output tokens, Terra — $2.50/$15, Sol — $5/$30. By third-party estimates, active daily work costs teams an average of $100–200 per developer per month; Plus with its base pool covers a moderate load.
Strengths and weaknesses in practice
Strengths:
- Parallelism. Cloud tasks don't occupy your machine: you can hand the agent three tasks and calmly work on a fourth.
- The GitHub flow. Codex lives in the familiar issue → branch → PR → review cycle; the
@codex reviewcommand in a pull request produces a structured code review. - Entry price. There's no separate subscription: $20 for Plus is the cheapest ticket into full agentic coding among the major players (as of July 2026).
- Openness. The CLI is open-source and supports MCP servers and the AGENTS.md format.
Weaknesses:
- Limits. On Plus, credits burn out fast on long tasks with the flagship model — you either watch your consumption or move to Pro.
- Dependence on the cloud and GitHub. Codex is at its strongest in cloud mode, which requires trusting your code to OpenAI's infrastructure and is built around GitHub.
- Quality on tangled legacy. Like any agent, Codex is confident on routine tasks and starts drifting on convoluted codebases with no clear instructions or tests.
- Access from Russia. Payment and account stability come with the caveats above; that's a risk factor if you need the tool for daily work.
- A younger customization ecosystem. Subagents, skills and hooks are further along in Claude Code — see the detailed comparison in Claude Code vs Codex.
Who Codex suits — and how not to marry the tool
Codex is a rational choice if you already pay for ChatGPT, live in the GitHub flow, and want a background performer for routine work: bug fixes, tests, refactors, small features. It also works for a beginner as a first agent, but it's worth understanding the basics first — what vibe coding is and how an agent differs from autocomplete.
If you're picking a tool, look at our ranking of vibe coding tools for 2026 and the overview of neural networks for writing code: the market turns over every six months, and today's best agent may well not be the best one a quarter from now.
Which leads to the main practical advice: don't tie your processes to a single tool. Project rules, policies and checklists belong in a portable layer. The qvib engine exports to the Claude Code and Cursor formats and to AGENTS.md — the very file Codex reads. Set up the discipline once, then try Codex, Claude Code or Cursor without rewriting the rules for each agent. We showed how that portability works in our breakdown of Claude Code vs Cursor vs VS Code.
FAQ
Is Codex free?
Formally yes: it's included even in the free ChatGPT plan (as of July 2026). But the limits there are token gestures and the model choice is trimmed. For real work you need at least Plus at $20/month, and for intensive use, Pro.
Does Codex only work with GitHub?
The cloud agent — yes, it's built around connected GitHub repositories. Codex CLI and the IDE extensions work with any local folder and any git host: GitLab, Bitbucket, self-hosted.
Which model is behind Codex?
As of July 2026 — the GPT-5.6 family: Sol (maximum quality), Terra (balance), Luna (speed and price). On paid plans the model and effort level are chosen manually; free plans use the entry-level configurations.
Can you use Codex from Russia?
The region is officially unsupported. The legal options come down to a foreign card issued in your name or intermediary services — with markups and risks. If you need a stable tool here and now, look at the alternatives in the roundup of neural networks for code above.
Does Codex replace a programmer?
No. Codex speeds up execution: routine work, tests, standard features. Framing the task, architecture, review and accountability for the result stay with the human — and the clearer your rules (AGENTS.md included), the more useful the agent is.