What you’re seeing is a glimpse into the near future: four Codex agents running in parallel, each building software I’ve assigned. There’s a new model in town—GPT‑5 Codex—and it’s changing how and where we code.
GPT‑5‑Codex is OpenAI’s newest coding model, tuned for “agentic” software work in Codex (terminal/IDE + cloud). It pairs with you for quick edits and can also run independently on long tasks (refactors, test‑driven bug fixes, code review). It’s available inside Codex today (default for cloud tasks & code review), not yet as a standalone API model. Install the CLI or IDE extension, connect GitHub for cloud, add an AGENTS.md to steer the agent, and keep internet access sandboxed/allow‑listed. (OpenAI)
GPT‑5‑Codex is a variant of GPT‑5 optimized specifically for coding inside Codex. It was trained on real engineering workflows (building features, debugging, large‑scale refactors, and PR reviews) and is designed to both pair interactively and execute longer tasks on its own. In OpenAI’s announcement, it’s now the default model for cloud tasks and code review in Codex; you can also opt into it locally via the CLI/IDE. (OpenAI)
OpenAI also published a system‑card addendum outlining additional safety training and product‑level mitigations (sandboxing, configurable network access). (OpenAI)
Availability note: GPT‑5‑Codex isn’t a public API model yet; you won’t see
gpt-5-codexin the API model list or ChatGPT model picker. It’s used inside Codex (web/CLI/IDE) and enabled by your ChatGPT plan. (OpenAI Help Center)
Codex (which uses GPT‑5‑Codex for cloud & review) is included in ChatGPT Plus, Pro, Business, Edu, and Enterprise. Local usage can also be billed via API key at standard API rates if you prefer usage‑based billing. See Codex pricing & limits and ChatGPT plan details:
Works on macOS & Linux (Windows via WSL). (OpenAI Developers)
# npm
npm install -g @openai/codex
# or Homebrew
brew install codex
codex # choose “Sign in with ChatGPT” for plan-based usage
/model gpt-5-codex medium # try "high" for complex tasks
(You can also choose older API models with --model, but GPT‑5‑Codex is recommended inside Codex.) (OpenAI Developers)
# Understand a repo
codex "Give me a high-level tour of this codebase; map key modules and list top 5 risks."
# Fix a failing test (paste logs)
codex "Use the stack trace below to find and fix the bug. Run tests until green."
# Multi-file refactor
codex "Thread a request_id through auth → handlers → logger; update tests and PR message."
(Attach images with -i if useful: codex -i screenshot.png "Explain this UI bug".) (OpenAI Developers)
Open‑source repo:https://github.com/openai/codex (Apache‑2.0). (GitHub)
Install the Codex IDE extension for VS Code/Cursor/Windsurf from the marketplace, sign in with your ChatGPT account, then choose GPT‑5‑Codex and set your reasoning level. Windows is supported experimentally (best via WSL). (OpenAI Developers)
Helpful tip: put Codex in the right sidebar and use the model/effort switcher directly under the chat input. (OpenAI Developers)
Docs: https://developers.openai.com/codex/ide (OpenAI Developers)
Example prompts for cloud:
<pkg>—find and fix; add tests; prepare a PR.” (OpenAI Developers)Network access is off by default; enable carefully with an allow‑list (e.g., package registries) and restrict HTTP methods to GET/HEAD/OPTIONS where possible. (OpenAI Developers)
Enable Code review in Codex settings for your repository, then mention @codex review on a PR to trigger a review. Codex analyzes the diff and codebase, runs tests to validate behavior, and posts comments in the PR. (OpenAI Developers)
Docs: https://developers.openai.com/codex/cloud/code-review (OpenAI Developers)
AGENTS.md (your agent’s playbook)Add an AGENTS.md at the repo root to encode build commands, test/lint rules, coding conventions, and PR expectations. Codex (and other agents) will read this and follow your house style and workflows—think of it as a README for agents. (Agents)
AGENTS.md to improve steerability. (OpenAI)Starter template you can paste:
# AGENTS.md
## Setup
- Install deps: `pnpm install`
- Run tests: `pnpm test`
- Lint & format: `pnpm lint && pnpm format`
## Style
- TS strict mode; prefer pure functions; single quotes, no semicolons.
## Verification
- CI must pass: `pnpm test`
- Include PR title `[feat|fix|chore]: <scope> – <summary>`
- Summarize risks + test coverage in PR body.
OpenAI’s Prompting guide for Codex recommends:
Reusable prompt shells
<feature> across <files/modules>. Follow AGENTS.md. Write tests that cover <cases>. Keep public API stable. When finished, summarize changes and open a PR.”<pattern> across the repo. Preserve behavior. Update docs and tests. Run pnpm lint && pnpm test until green. Provide a migration note.”git diff in the chat before opening PR.”gpt-5-codex directly in the API; use Codex clients. (OpenAI Help Center)A) Local pairing (CLI)
# 1) Install & sign in
npm i -g @openai/codex && codex
# 2) Switch model & effort
/model gpt-5-codex high
# 3) Drive a refactor
"Refactor auth middleware to thread request_id; update tests; run until green."
# 4) Review & commit
"Summarize changes, risks, and test coverage. Prepare a conventional-commit message."
B) Cloud task (web/IDE)
/users, update API docs, and include unit/integration tests. Open a PR.” (OpenAI Developers)C) Code review (GitHub)
Comment on a PR: @codex review
Optionally: @codex review for security vulnerabilities or …for outdated dependencies. (OpenAI Developers)
Is GPT‑5‑Codex available via the API?
Not yet (as of 09/16/25). Use Codex (web/CLI/IDE). The Help Center calls out that gpt-5-codex isn’t currently an API model. (OpenAI Help Center)
Where do I turn it on?
In Codex: it’s the default in cloud & code review, and you can select it locally in CLI/IDE via the model switcher. (OpenAI)
Is it safe for sensitive repos?
Codex runs in a sandbox by default; keep internet access off or tightly allow‑listed and review changes. See the safety docs and system‑card addendum. (OpenAI Developers)