Codex
Works through the list
Codex is OpenAI's coding agent, across the ChatGPT apps, the terminal, your IDE and the cloud.
- Cost
- Free to startChatGPT Plus is US$20 a month
- Setup, roughly
- 10 minutes
- Runs on
- App, terminal, code editor or cloud
- Works on
- Mac, Windows or Linux
Checked 7 Nov 2026.
You Read this repo and explain how it's put together. Then suggest the smallest change that would fix this problem. Don't make changes yet.
Found it. The bug is in the date parser.
parse.ts
- const d = new Date(input)
+ const d = parseISO(input)
+ if (!isValid(d)) return nullYour tests cover this path. Nothing changed yet
Our take
A strong choice for building software, especially where a command can check the work.
Our stake: Navigate helps clients set up agents like this one. Claude helps research and draft this site, so a person reviews this page. Nobody paid for this guide.
What is Codex best at?
- Code reviewCloud review on pull requests.
- Cloud tasksHand off work and come back later.
- Any surfaceApp, terminal, IDE and phone.
- Checked workBest when a command can verify it.
See Codex at work
OpenAI's own video, on YouTube. Nothing loads from YouTube until you press play.
How much does Codex cost?
With only an API key, cloud features such as code review and Slack are not included. Prices are in US dollars. Checked 7 Nov 2026.
How do I set up Codex safely?
6 steps, about 10 minutes today, then a week of checking. We put the approvals on before anything else.
- Pick where to use it
- Protect your work first
- Start in a real project
- Ask before it changes anything
- Give it a way to check
- End in a pull request
See every step
1. Pick where to use it. The ChatGPT app is the gentlest start. The CLI and code editor extension suit developers.
2. Protect your work first. Work in a project tracked in Git, with everything committed, so you can undo any change. Keep its default approval mode, and don't switch to full access or any mode that skips approvals.
3. Start in a real project. Point it at one repo you know well.
4. Ask before it changes anything. Start by asking it to explain the codebase.
5. Give it a way to check. Tell it which command proves the work is right.
6. End in a pull request. Keep a person on every merge.
Is Codex safe?
We would use it when a person reviews every change. This is what we would watch.
- End in a pull requestProtect your main branch so a person has to merge. Don't rely on the agent to hold back.
- Limits are rangesCheck your usage each week.
- Keep it updatedSecurity fixes land often. Update it before you open code you didn't write.
- Keep money keys outNo bank or ad-account keys anywhere it can read.
Who should skip Codex?
- There is no check it can run
- You only have an API key but want cloud review, Slack or cloud tasks
How has Codex done in tests?
Coding agents have real, repeatable tests. See who tests what.
- Personal agentsPersonalAgentBench, by micro1. Not tested
- Coding agentsAA Coding Agent Index, by Artificial Analysis. Tests this product
- Self-hosted agentsWildClawBench, by InternLM. Tests this product
- Office tasksGDPval-AA, by Artificial Analysis. Tests its model, not the product
- Running a businessVending-Bench, by Andon Labs. Tests its model, not the product
- Freelance workRemote Labor Index, by Center for AI Safety and Scale Labs. Tests this product
Alternatives to Codex
- Claude CodeDevelopers and technical teams. From US$20 a month.
The details
Good to know
Where it runs. ChatGPT apps, the CLI, the code editor (IDE) extension and Codex Cloud.
With an API key. CLI, SDK or code editor extension, billed at API rates.
Sources for this page
Pricing (OpenAI docs), read 7 Nov 2026
Quickstart (OpenAI docs), read 7 Nov 2026
Know better?
Tell us what we've missed at info@navigateagents.com. We'll fix it and credit you if you'd like.
