OpenAI Codex Explained: What It Is and How It Works
A plain-language guide to OpenAI's agentic coding system for business and engineering teams.
OpenAI Codex is an agentic coding system that turns a plain-language request into real code changes. You describe a task. Codex reads your codebase, edits files, runs commands, and pauses for your review.
It is not a single app. Codex is an umbrella over several surfaces that share one model and one account: a terminal CLI, an IDE extension, cloud tasks in ChatGPT, and a GitHub bot.
This page explains what Codex is, which model runs it, what it costs, and where it fits for business teams. It links to deeper guides on pricing, hardware, and head-to-head comparisons.
What Is OpenAI Codex?
OpenAI Codex is an autonomous coding agent that writes and changes code from natural-language instructions. You give it a goal, and it does the work across your files.
The difference from a chat window is autonomy. Codex does not just suggest a snippet. It searches your repository, makes edits, runs your tests, and reports back.
It can run short tasks or work for hours on a larger job. Longer sessions use isolated worktrees and cloud environments so your main branch stays clean.
Think of Codex as a junior engineer who takes a ticket, opens a branch, and comes back with a reviewable change. You stay in control at the review step.
- Reads and searches your whole codebase, not one file
- Edits files and runs commands or tests on its own
- Pauses for human review before changes are final
- Handles both quick fixes and multi-hour autonomous work
Deciding whether Codex belongs in your engineering stack? Layer3 Labs runs a hands-on workflow audit and gives you a clear adoption plan.
Book a ConsultationThe Codex Surfaces: One Model, One Account
Codex ships as several surfaces that all share one model and one account context. Where you start a task does not change the underlying engine.
The terminal CLI installs from npm and runs on macOS, Linux, and Windows through PowerShell with WSL. It supports /plan, /exec, and /review commands plus worktrees and cloud environments.
The IDE extension brings the same agent into your editor. Cloud task delegation lets you hand off jobs from ChatGPT and collect the results later.
A GitHub bot reviews pull requests and can act on your repository. Codex can also use computer-use by reading the screen. This shared design means context follows you across every surface.
- Terminal CLI (npm; macOS, Linux, Windows via PowerShell + WSL)
- IDE extension inside your editor
- Cloud task delegation through ChatGPT
- GitHub bot and automated code review
- Computer-use via screen reading
Which Model Powers Codex?
Codex runs on the GPT-5.6 model family from OpenAI. That family includes three tiers named Sol, Terra, and Luna.
Sol is the most capable tier, and Sol Ultra is rolling into the Codex client. Terra and Luna trade some capability for speed and lower cost on lighter tasks.
You do not usually pick a model by hand. Codex routes work to the right tier based on the job and your plan.
This matters for regulated teams. Confirm the exact model in your agreement before you point Codex at sensitive code.
- Model family: GPT-5.6 (Sol, Terra, Luna)
- Sol Ultra is rolling into the Codex client
- Codex routes tasks to the right tier automatically
How to Use Codex: Task to Edits to Review
You use Codex by describing a task in plain language, then approving its work. The loop is simple and repeats until the job is done.
First, you write the goal, such as "add input validation to the signup form." Codex searches your code to understand the context.
Next, it edits the files it needs and runs your commands or tests. If something fails, it tries again and keeps iterating.
Finally, it stops and shows you a reviewable change. You approve, ask for edits, or reject it. Nothing ships without your sign-off.
- Describe the task in plain English
- Codex reads the codebase and plans the change
- It edits files and runs tests to verify
- It pauses so you can review before merging
What Plans Include Codex and What It Costs
Codex is included in every ChatGPT plan, from Free to Enterprise. The plan sets how much usage you get, not whether you get access.
Paid tiers run from Go at $8 per month up to Pro at $100 per month, with a Pro 20x option at $200. Business is about $25 per user monthly, and Enterprise is custom with a shared credit pool.
On April 2, 2026, OpenAI moved Codex usage to API-token-aligned credits instead of per-message limits. Heavier tasks now draw more credits.
Real-world spend lands near $100 to $200 per developer per month, depending on how hard the team leans on it. See the full breakdown on our pricing guide.
- Free: $0
- Go: $8/mo
- Plus: $20/mo
- Pro: $100/mo (Pro 20x: $200/mo)
- Business: ~$25/user/mo
- Enterprise: custom, shared credit pool
Codex Micro: The New Hardware Keypad
Codex Micro is a $230 physical keypad, the first hardware accessory built for Codex. It gives you one-press access to common agent actions.
The keypad is made in partnership with Work Louder. It sits next to your keyboard and maps keys to Codex commands like plan, run, and review.
It is optional. Codex works fully without it, but heavy users may like the tactile shortcuts.
For a full look at the device and whether it earns desk space, see our dedicated guide.
Best Uses for Business Teams
Codex pays off most on repetitive, well-scoped engineering work. These are tasks with clear inputs and testable outputs.
Good fits include writing tests, fixing bugs from a clear ticket, and routine refactors. It also handles dependency bumps and boilerplate scaffolding well.
It is weaker on vague, high-stakes design decisions. Those still need a human to set direction first.
The teams that win treat Codex as leverage on defined work, not a replacement for judgment. Clear tickets get clear results.
- Writing and expanding test coverage
- Fixing well-described bugs
- Routine refactors and code cleanup
- Dependency updates and boilerplate scaffolding
Does Codex Train on Your Code? Data and Compliance
OpenAI does not train on ChatGPT Business, Enterprise, or API workspace data by default. Your code in those tiers is not used to improve the model.
The platform carries SOC 2 and ISO 27001 coverage, which most security reviews expect. That helps with vendor approval.
For regulated data, confirm the specific model in your BAA before you connect it. Default protections are strong, but named agreements remove doubt.
The tradeoff to note: default free and personal tiers do not carry the same guarantees. Put regulated work on Business or Enterprise, not personal accounts.
- No training on Business, Enterprise, or API data by default
- SOC 2 and ISO 27001 platform coverage
- Confirm the exact model in your BAA for regulated data
- Keep sensitive work off free and personal tiers
How Codex Compares to Claude Code and Cursor
Codex, Claude Code, and Cursor are the three leading agentic coding tools, and each has a different center of gravity. Codex is strongest when your team already lives in the OpenAI and ChatGPT ecosystem.
Claude Code is a terminal-first agent from Anthropic that many teams pick for careful, review-heavy work. Cursor is an AI-native editor that wraps multiple models inside a full IDE.
The honest split: Codex leads on bundled ChatGPT access and cloud delegation, Claude Code on deep terminal control, and Cursor on an all-in-one editing experience.
Pick based on where your team already works. Read the full head-to-head breakdowns before you standardize.
Limitations to Plan Around
Codex still needs human review, so it is not a hands-off replacement for engineers. Its edits can be wrong or incomplete on complex work.
Costs scale with usage under the token-aligned credit model. Heavy autonomous sessions can run up spend faster than a flat per-seat tool.
It works best with clear tickets and good tests. Vague goals or weak test coverage lead to weaker results.
Plan for a review process and usage limits before you roll it out team-wide. That keeps quality and budget in check.
- Output needs human review before it ships
- Token-based credits mean cost scales with heavy use
- Results depend on clear tasks and strong tests
- Team rollout needs review and budget guardrails
Frequently Asked Questions
- Yes, Codex is included in the free ChatGPT plan at $0. Every paid tier also includes it, from Go at $8 per month up to Pro at $100 and custom Enterprise. Your plan sets your usage limits, not your access.
- Codex runs on the GPT-5.6 family from OpenAI, which includes the Sol, Terra, and Luna tiers. Sol is the most capable, and Sol Ultra is rolling into the Codex client. Codex routes each task to the right tier automatically.
- No, Codex is a coding agent, while ChatGPT is a chat assistant, though they share one account. ChatGPT answers questions in a chat window. Codex reads your codebase, edits files, and runs tests, then pauses for your review.
- Yes, Codex has a terminal CLI installed from npm. It runs on macOS, Linux, and Windows through PowerShell with WSL. It supports /plan, /exec, and /review commands plus worktrees and cloud environments.
- No, OpenAI does not train on ChatGPT Business, Enterprise, or API workspace data by default. The platform carries SOC 2 and ISO 27001 coverage. For regulated data, confirm the exact model in your BAA before connecting it.
- Codex can search your codebase, edit files, run commands and tests, and iterate until a task is done. It can work for multiple hours in isolated worktrees. It still pauses for human review before any change is final.
Bring Codex Into Your Team the Right Way
Layer3 Labs helps teams adopt agentic coding tools like Codex without the false starts. We map the surfaces to your workflow, set guardrails for cost and review, and train your engineers to get real leverage.
Book a Consultation