It is for developers who want an agent to read a repo, edit code and run commands under explicit sandbox and approval controls, and to script it in CI with codex exec. Recent changes center on the GPT-6.1 Sol model replacing GPT-5.5 for ChatGPT sign-in, the retirement of the untrusted approval policy and --full-auto, and CLI 0.164.0 adding /presets and inline /model and /permissions arguments. Codex usage comes with ChatGPT plans from Plus up, or is billed at API rates with an API key.
*Field guide compiled 2026-10-23 from the official documentation, pricing pages and changelogs linked below; every line links to the page that states it.*
What it does
- Codex CLI. Lets Codex inspect files, make edits and run tools already installed on your machine from the terminal, with your choice of model, reasoning effort, permissions and commands. How: Install on macOS/Linux with curl -fsSL https://chatgpt.com/codex/install.sh | sh, sign in, and run codex from a project directory. source
- Sandbox mode. Controls what Codex can technically do when it executes model-generated commands: read-only, workspace-write (the default for CLI and IDE: writes limited to the active workspace, no network) or danger-full-access. OS-level mechanisms enforce it. How: --sandbox read-only|workspace-write|danger-full-access, or sandboxmode in config.toml; switch to read-only mid-session with /permissions. source
- Approval policy. Controls when Codex must ask before acting, for example leaving the sandbox or using the network. The Auto preset is workspace-write plus on-request approvals; a granular policy keeps selected prompt categories interactive, and approvalsreviewer=autoreview routes eligible requests to Auto-review. How: --ask-for-approval on-request|never (-a never), or approvalpolicy in config.toml; --dangerously-bypass-approvals-and-sandbox (alias --yolo) removes both layers. source
- AGENTS.md. Instruction files Codex reads before doing any work: a global file in ~/.codex plus files from the project root down to the current directory, concatenated so closer files override earlier guidance. How: Create ~/.codex/AGENTS.md for personal defaults and AGENTS.md in the repo (AGENTS.override.md for temporary overrides); the combined size is capped by projectdocmaxbytes, 32 KiB by default. source
- codex exec. Non-interactive mode for scripts and CI: streams progress to stderr and prints only the final agent message to stdout, accepts piped stdin as context, and can emit a JSON Lines event stream or a final response matching a JSON Schema. How: codex exec "task"; add --json for JSONL events, --output-schema ./schema.json -o out.json for structured output, --ephemeral to skip saving session files, and --sandbox workspace-write to allow edits. source
- Codex IDE extension. Works beside your code with open files and selections as context, shows diffs in place, and can hand longer work to Codex cloud. How: Install the openai.chatgpt extension in VS Code, Cursor or Windsurf and choose the Codex icon (or run Codex: Open Codex Sidebar); in JetBrains open AI Chat and select Codex; in Xcode choose Codex as the agent. source
- Codex Cloud. Runs coding tasks in the cloud from a published environment; each task gets its own workspace and keeps working while your computer is asleep, and you review changes, request follow-ups, commit or open a pull request. How: In ChatGPT on web or desktop choose Work in Cloud, create or select an environment, connect GitHub repositories, publish, then start a task. source
- GPT-6.1 Sol. The recommended model for complex coding and agentic work, with near-Astra performance at lower cost; GPT-6 Luna is recommended for focused, repeatable tasks. Max gives more reasoning time, and Ultra splits work across subagents. How: codex --model gpt-6.1-sol, codex exec -m gpt-6.1-sol, /model in the CLI, or model = "gpt-6.1-sol" in config.toml. source
Modes
- Auto (preset). --sandbox workspace-write --ask-for-approval on-request; reads, edits and runs commands in the workspace, asks to edit outside it or use the network. source
- Safe read-only browsing. --sandbox read-only --ask-for-approval on-request. source
- Read-only non-interactive (CI). --sandbox read-only --ask-for-approval never; never asks for approval. source
- Auto-review mode. Same boundary as on-request, but eligible approval requests go to Auto-review (approvalsreviewer = "autoreview"). source
- Dangerous full access. --dangerously-bypass-approvals-and-sandbox (alias --yolo): no sandbox, no approvals, not recommended. source
Pricing and limits
- Free: $0/month. GPT-6 Luna at Standard speed in the desktop app, subject to rollout source
- Go: $8/month. GPT-6 Luna at Standard speed in the desktop app, subject to rollout source
- Plus: $20/month. Codex on the web, CLI, IDE extension and iOS; GPT-6.1 Sol and GPT-6 Luna; cloud integrations like automatic code review and Slack; extendable with ChatGPT credits source
- Pro: From $100/month (plans at $100, $200, or $500 USD per month). Everything in Plus; Ultrafast mode access on Pro $500 source
- Business: $20 / user / month (2+ users, billed annually. $25 per user per month when billed monthly.). Larger virtual machines for cloud chats, admin controls, SAML SSO, no training on business data by default source
Recent changes
- 2026-10-23: Codex CLI 0.164.0: ten numbered /presets slots for reusable prompts and slash commands, inline /model and /permissions arguments, image previews on Kitty or Sixel terminals, enterprise MCP authentication, and resumed sessions restoring saved thread settings. source
- 2026-10-14: GPT-5.5 retired from Codex for users signed in with ChatGPT; use GPT-6.1 Sol where available. source
- 2026-10-09: Composer predictions beta suggests a follow-up prompt in the Codex desktop app; press Tab to accept as an editable draft. source
- 2026-10-08: GPT-6.1 Sol Ultrafast mode launched in Codex and ChatGPT Work for Pro $500 and eligible Enterprise and Edu plans. source
- 2026-09-29: GPT-6.1 Sol added to Codex and ChatGPT Work. source
Traps
- --full-auto is no longer supported in codex exec; scripts must switch to --sandbox workspace-write or --approve-for-me. source
- codex exec runs in a read-only sandbox by default, so CI jobs that need edits must pass --sandbox workspace-write. source
- approvalpolicy = "untrusted" is no longer supported; use sandboxmode = "read-only" with approvalpolicy = "on-request" instead. source
- In the default workspace-write sandbox, network access is off unless enabled in configuration, and writableroot/.git stays read-only. source
- AGENTS.md files stop being added once the combined size reaches projectdocmaxbytes (32 KiB by default). source
- GPT-5.5 retired from Codex for ChatGPT sign-in on October 14, 2026; saved settings, managed configs, custom agents and scripts that select gpt-5.5 need updating. API-key sessions are unaffected. source