🇺🇸 USA · OpenAI Codex (2026 agent)
Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: OpenAI’s coding-agent answer to Claude Code — released 2025, generally available 2026. A cloud-and-local agent that takes a task, plans it, writes code, runs tests, and commits results.
Important name disambiguation: “Codex” has been used by OpenAI for TWO different products. The original OpenAI Codex (2021-22) was the underlying model that powered the first GitHub Copilot — deprecated. The 2025/2026 Codex covered in this entry is the new coding agent product announced May 2025 (cloud agent) plus the Codex CLI (terminal agent) — completely separate product line, just sharing the name.
Front-matter facts
| Field | Value |
|---|---|
| Vendor | OpenAI (San Francisco, USA) |
| Country / origin | 🇺🇸 USA |
| Recommended for Australian users? | ✅ Yes — fully accessible from AUS via ChatGPT and OpenAI API |
| Privacy summary | ChatGPT Plus / Pro tiers: opt-out of training applies to Codex inputs. Team / Enterprise: never trains. Codex CLI on API: never trains by default. |
| Free tier | Limited — basic Codex tasks available in ChatGPT Plus; full agent access in Pro |
| Paid tiers | Available with ChatGPT Plus US200/mo (heavy Codex use), ChatGPT Team / Enterprise (codex with no-training), OpenAI API for Codex CLI (pay-per-token) |
| First released | New Codex agent announced May 2025; generally available 2026 |
| Last reviewed | 2026-06-26 |
| Official site | https://openai.com/codex (and accessed via chatgpt.com / Codex CLI) |
What it is
The new OpenAI Codex (2025/26) is OpenAI’s coding agent product — its strategic answer to Anthropic’s Claude Code. It comes in two forms:
- Codex in ChatGPT (cloud agent) — accessed inside ChatGPT at chatgpt.com/codex; runs in cloud sandboxes. You give it a task (“add OAuth login to this repo”); it spins up an isolated environment, clones your repo, plans the work, writes code, runs tests, opens a pull request.
- Codex CLI — a terminal-based agent (similar in shape to Claude Code) that runs locally on your machine, modifying files in your local repo. Open-source CLI, talks to the OpenAI API.
It’s powered by codex-1 (a GPT-5-family model fine-tuned specifically for agentic coding) — distinct from the general GPT-5 used in standard ChatGPT.
Codex CLI is open-source, hosted on GitHub at openai/codex. The cloud agent is closed-source and runs on OpenAI’s infrastructure.
Why this matters: until 2025, OpenAI had no direct competitor to Claude Code — Codex was the original 2021 model, deprecated. GitHub Copilot was the closest OpenAI-adjacent product but doesn’t have the agent depth Claude Code introduced. The 2025 Codex relaunch closed that gap. Each major frontier lab now has its own coding-agent product:
- Anthropic Claude Code (terminal-first, integrated with Claude API + Skills + MCP)
- OpenAI Codex (cloud-first + CLI option)
- Google Antigravity (Google’s agentic IDE — separate entry)
What you’d use it for
Codex in ChatGPT (cloud agent)
- “Add feature X to my repository” — Codex clones repo in cloud sandbox, implements, runs tests, opens PR
- Long-running tasks (10-60+ minutes) without occupying your terminal
- Multi-step refactoring across many files
- Background tasks that you can check on while doing other work
Codex CLI
- Same use cases as Claude Code (terminal-first agentic coding)
- Local file modification
- Integrated with your local terminal, git, environment
- Open-source — auditable, customisable, self-contained
How to sign up + first 5 minutes from Australia
Codex in ChatGPT
- Subscribe to ChatGPT Plus (US200/mo) at chatgpt.com.
- Go to
chatgpt.com/codex(or click the Codex link in the ChatGPT sidebar). - Connect your GitHub account when prompted (authorise OpenAI to read/write specific repositories).
- Give Codex a task: “In repo geo-au/foo, add an /about page using Next.js.”
- Codex clones, plans, codes, tests, opens a PR. You review the PR in GitHub before merging.
Codex CLI
- Install:
npm install -g @openai/codex(or via Homebrew:brew install codex) - Set your OpenAI API key:
export OPENAI_API_KEY=sk-...(sign up at platform.openai.com for an API key first) - Open a terminal in your project directory; run
codex - Codex CLI launches in interactive mode — you can chat, give tasks, run autonomous steps
- CLI runs locally; modifies files in your local repo
What it costs — what you actually get
Cloud Codex
- ChatGPT Plus US$20/mo — limited cloud Codex tasks per month
- ChatGPT Pro US$200/mo — heavy cloud Codex usage, longer task runtimes, priority queue
- ChatGPT Team US$25-30/seat/mo — team-shared Codex; no training on code
- ChatGPT Enterprise — high limits, custom contracts
Codex CLI (open-source)
- The CLI itself is free (open-source MIT license)
- API token usage is billed via OpenAI API (pay-per-token to platform.openai.com)
- Typical cost for a moderate coding session: US$1-10 in API calls (varies dramatically by model used and task length)
Hidden costs to know about
- Cloud Codex usage is bundled in ChatGPT Plus/Pro/Team — there’s no separate metering for individual cloud Codex tasks at consumer tier
- Codex CLI on OpenAI API is metered per token — heavy use can add up (similar to Claude Code via Anthropic API)
- For predictable monthly costs at heavy use, prefer ChatGPT Pro (cloud) over Codex CLI (API)
How it compares to alternatives
| Capability | OpenAI Codex (cloud) | OpenAI Codex CLI | Claude Code | Google Antigravity |
|---|---|---|---|---|
| Surface | Cloud in chatgpt.com | Local CLI | Local CLI | IDE-style cloud agent |
| Underlying model | codex-1 (GPT-5 family) | codex-1 + general GPT-5 | Claude (Opus / Sonnet / Haiku / Fable) | Gemini |
| Open-source | No | Yes | No (CLI is closed) | No |
| Long background tasks | Excellent (runs in cloud) | Local-only | Local + Remote Desktop mode | Yes |
| GitHub PR integration | Native | Manual | Manual | Native |
| Skills / Plugins system | Limited | Limited | Rich (Claude Skills + Plugins) | Limited |
| MCP support | Newer | Yes | Native (best) | Yes |
| Subscription bundling | ChatGPT Plus / Pro | API | Claude Pro / Max | Google AI Pro / Ultra |
| Cost (heavy use) | ChatGPT Pro US$200/mo | API metered (variable) | Claude Max US$100-200/mo | Google AI Ultra US$249/mo |
The honest take:
- For cloud-agent “fire and forget” tasks, Codex in ChatGPT is currently best-in-class
- For terminal-first daily-driver coding, Claude Code is the encyclopedia’s primary recommendation (richer ecosystem, Skills, Plugins, MCP)
- For open-source / auditable local agent, Codex CLI is genuinely open and a strong choice
- Many developers use both — Claude Code locally for active development + Codex cloud for “while I’m at lunch, refactor this”
Privacy / data handling
ChatGPT Plus / Pro tier Codex: subject to same training-on-by-default policy as general ChatGPT. Opt-out via Settings → Data Controls → Improve the model for everyone → OFF.
ChatGPT Team / Enterprise Codex: does NOT train on your code.
Codex CLI: uses OpenAI API; API does NOT train by default (since March 2023 commitment).
GitHub integration: Cloud Codex receives explicit per-repository authorisation; you can revoke at any time in GitHub Settings → Applications.
Sandbox isolation: Cloud Codex tasks run in ephemeral sandboxes that are destroyed after task completion.
Where data lives: US data centres primarily. Enterprise can negotiate data-residency.
Recent changes
- 2026: Codex CLI generally available; integrated with OpenAI Agents SDK
- 2025: New Codex announced (May 2025); cloud Codex launched in ChatGPT
- 2022-23: Original Codex deprecated as standalone model
- 2021: Original OpenAI Codex model launched (powered first GitHub Copilot)
(Check openai.com/blog for the latest.)
Gotchas
- “Codex” overloaded name is the biggest confusion. 2021 Codex (deprecated model) ≠ 2025 Codex (new agent product). When someone says “Codex” check which they mean.
- Cloud Codex requires GitHub authorisation per repo — review what you’re granting access to.
- Cloud Codex sandboxes have limits — long-running tasks (multi-hour) may time out; consider breaking into smaller chunks.
- Codex CLI uses API billing, not ChatGPT subscription — easy to rack up API costs if you assume your Plus subscription covers it. It doesn’t.
- Codex’s open-source CLI is genuinely auditable — the source on github.com/openai/codex can be inspected. Unusual for OpenAI to open-source production tooling.
- Skills / MCP support is newer than Claude Code’s — for the deepest ecosystem of skills + connectors, Claude Code remains ahead in mid-2026.
- PR integration is native to GitHub — for GitLab / Bitbucket, the integration is weaker.
See also
- Claude Code deep dive 🟩 🟦 — primary alternative
- Cursor 🟩 🟦
- Windsurf 🟩 🟦
- GitHub Copilot 🟩 🟦 — separate OpenAI-adjacent product (different SKU)
- Google Antigravity 🟥
- OpenAI API 🟥 — Codex CLI’s billing path
- ChatGPT 🟩 🟦
- OpenAI Agents SDK 🟥
- MCP — Model Context Protocol 🟩 🟦
- Decision frameworks — Cursor vs Claude Code vs Windsurf 🟥