OpenAI Codex is the hardest name in developer tooling to write about accurately, because it has denoted three genuinely different products since 2021 and most explanations quietly blend them.
Today it is a coding agent product line spanning a desktop app, a command line tool, IDE extensions, a cloud service, GitHub and Chrome. It is not a model. Understanding that, and the two things the name used to mean, is most of what you need before evaluating it.
OpenAI Codex has named three different products
2021 to 2023: Codex was a model. A GPT-3 descendant fine-tuned on code, available through the API as code-davinci-002 and code-cushman-002, and the engine behind the original GitHub Copilot. OpenAI announced its deprecation on March 20, 2023 and shut the models down on March 23, a three-day window that was unusually abrupt and is still remembered for it.
2025: Codex was a cloud agent. Relaunched on May 16, 2025 as, in OpenAI’s words, "a cloud-based software engineering agent that can work on many tasks in parallel, powered by codex-1," itself described as "a version of OpenAI o3 optimized for software engineering." It ran entirely inside isolated cloud containers, available first to Pro, Business and Enterprise subscribers.
Today: Codex is a product line. OpenAI’s developer documentation frames it as "OpenAI’s coding agent for software development," an umbrella across many surfaces.
Nothing technical survives from the 2021 model. Only the brand does. If a comparison you are reading discusses Codex’s benchmark scores against Copilot, check which decade it means.
What you actually get today
The surfaces, with their real status:
- Desktop app, generally available, macOS since February 2, 2026 and Windows since March 4, 2026
- CLI, generally available, Apache-licensed and open source, installable through npm, Homebrew or a shell script
- IDE extension, generally available, for VS Code, Cursor and Windsurf, plus JetBrains editors
- Cloud and web, generally available, for background and parallel tasks
- GitHub, generally available: tag the agent on an issue or pull request, and it opens pull requests, performs automated code review and holds conversations in the pull request
- Mobile, generally available, but as a remote control of a desktop Codex rather than a local runtime
- Chrome extension, since May 7, 2026, working across tabs in the background
- Sites, which builds and hosts sites from the app, still in preview
- Amazon Bedrock, generally available since June 1, 2026
The mobile distinction is the one that trips people. You are not running an agent on your phone; you are steering one running on your machine, which we covered when it launched in Codex on mobile.
The OpenAI Codex sandbox and approval model
This is the best-documented part of Codex and the most useful thing to understand before pointing it at a repository.
The security documentation is unusually specific. There are three sandbox modes: read-only, which is the default in folders not under version control; workspace-write, the default in folders that are; and danger-full-access, which disables the protections. Approval policies are separate and run untrusted, on-request and never, with a granular table form for finer control.
Enforcement is real and per-platform. macOS uses Seatbelt through sandbox-exec with a profile matching the selected mode. Linux uses bubblewrap plus seccomp. Windows uses WSL2 where available and a native sandbox otherwise.
The network default is the detail that saves debugging time. Locally, "the agent runs with network access turned off" by default. In the cloud, Codex uses a two-phase runtime: a setup phase with network access to install dependencies, then an agent phase that runs offline by default unless you explicitly enable internet access for that environment. If your build fails because a package could not be fetched, this is why.
Which models run it, and the wrinkle
Codex today runs on general purpose GPT models rather than Codex-branded ones. The picker offers GPT-5.5, which the documentation recommends starting with, plus GPT-5.4, GPT-5.4-mini, GPT-5.3-codex and a research-preview GPT-5.3-codex-spark for near-instant iteration. You select with -m on the command line, /model inside a thread, or the IDE selector, and reasoning effort is adjustable.
Here the picture genuinely contradicts itself, and it is worth stating rather than smoothing. A changelog entry from May 26, 2026 says GPT-5.3-Codex and GPT-5.2 were discontinued as user-selectable options for ChatGPT-signed-in users, with API keys unaffected, yet the models and pricing pages still list GPT-5.3-Codex. Separately, the API deprecations page retired the whole gpt-5-codex through gpt-5.2-codex line on July 23, 2026 in favor of gpt-5.6-sol and gpt-5.6-terra.
The direction is clear enough even if the details are messy: the dedicated Codex model line is being folded into the flagship line. Two more dates matter to anyone planning around this. GPT-5.4 and GPT-5.4-mini leave Codex on August 31, 2026, announced on July 31.
Pricing and the published limits
Codex is included with Plus, Pro, Business, Edu and Enterprise plans. Free and Go have time-limited access. Plus is $20 a month, Go is $8, and Pro starts at $100.
OpenAI publishes usage limits as ranges per five-hour window, which is unusual and more honest than a single number. On Plus you get 15 to 80 local messages on GPT-5.5, 20 to 100 on GPT-5.4, and 30 to 150 local plus 10 to 60 cloud tasks on GPT-5.3-Codex. Pro at the 5x tier multiplies those roughly fivefold, and the 20x tier by about twenty.
Past the included limits, credits extend usage, and billing moved from per-message to token-based on April 2, 2026. OpenAI notes that a typical Codex task on GPT-5.5 consumes between 5 and 45 credits.
If you are weighing Codex against the alternative, our comparison of Claude Code and OpenAI Codex covers the head-to-head, and our look at ChatGPT’s split into Chat, Work and Codex covers where Codex sits inside OpenAI’s wider product reshuffle.
Configuring it
Configuration lives in ~/.codex/config.toml, in TOML rather than JSON or YAML. A project-scoped .codex/config.toml loads only when you have marked the project as trusted, which is a sensible default given that a config file can alter approval behavior. Named profiles let you switch between setups with a flag.
MCP servers are supported in both the CLI and the IDE extension with shared configuration, over stdio and streamable HTTP with bearer token or OAuth. You can enable and disable individual tools, set default approval modes per server, and configure timeouts. Add servers with codex mcp add and inspect them with /mcp.
AGENTS.md is the project instruction convention, and the /init command scaffolds one. Worth flagging honestly: the canonical documentation page for AGENTS.md was not resolvable at the time of writing, so treat any specific claim about precedence or merge order as unverified until OpenAI republishes it.
What changed recently
The pace here is the thing to plan around. Since May 2026: access tokens for enterprise automation and the Chrome extension in May; Bedrock support, the Sites preview, migration flows from Claude Code and Cowork, rate-limit reset banking and /init in June; task management from conversations, Codex folding into the ChatGPT desktop app, pull request chat, custom domains for Sites, voice input, multi-folder projects, a Sign in with ChatGPT beta across Airtable, GitLab, HubSpot, Notion, Supabase and Vercel, and multi-repo code review in July.
That is roughly one meaningful change a week, which is the single most important thing to know about OpenAI Codex as a purchasing decision. Any version number in this post is stale by the time you read it, which is why the changelog is a better reference than any article, including this one.
Frequently Asked Questions
What is OpenAI Codex?
Today it is OpenAI’s coding agent product line, spanning a desktop app, a command line tool, IDE extensions, a cloud service, GitHub integration, a Chrome extension and mobile remote control. It is a product rather than a model, and it runs on general purpose GPT models. The name previously belonged to a code completion model from 2021 and to a cloud software engineering agent launched in 2025.
Is Codex the same thing that powered GitHub Copilot?
No, not any more. The original Codex models did power the first version of Copilot, but OpenAI announced their deprecation on March 20, 2023 and shut them down three days later. Nothing technical carries over to the current product. Any comparison citing Codex model benchmarks against Copilot is describing a system that has not existed for over three years.
Which models does Codex run on?
General purpose models rather than Codex-branded ones. The picker currently offers GPT-5.5, which the documentation recommends as the starting point, along with GPT-5.4, GPT-5.4-mini, GPT-5.3-codex and a research-preview spark variant. GPT-5.4 and GPT-5.4-mini leave Codex on August 31, 2026, and the dedicated Codex model line is being folded into the flagship GPT-5.6 family.
Is Codex included in my ChatGPT subscription?
It is included with Plus, Pro, Business, Edu and Enterprise. Free and Go have time-limited access. Usage limits are published as ranges per five-hour window rather than fixed numbers, so Plus gives 15 to 80 local messages on GPT-5.5 with the Pro tiers multiplying that roughly five or twenty times. Credits extend usage past the included allowance.
Does Codex have internet access when it runs?
Not by default, in either place. Running locally, the agent starts with network access turned off. In the cloud it uses a two-phase runtime where a setup phase can reach the network to install dependencies and the agent phase then runs offline unless you explicitly enable internet access for that environment. This is the most common cause of unexpected build failures.
How does the Codex sandbox work?
Three sandbox modes control filesystem access: read-only, which is the default outside version control, workspace-write, the default inside it, and a full-access mode that disables protections. Approval policies are configured separately. Enforcement is per-platform, using Seatbelt on macOS, bubblewrap with seccomp on Linux, and WSL2 or a native sandbox on Windows.
Can I run Codex on my phone?
Only as a remote control. The ChatGPT iOS app can steer a Codex instance running on your desktop, review its output and approve steps, but the agent itself does not run locally on the device. This is an architectural choice rather than a temporary limitation, and it means your desktop needs to be awake and running for mobile use to work.
How do I configure Codex for a project?
Global configuration lives in a TOML file at ~/.codex/config.toml, with project-scoped configuration in .codex/config.toml that loads only for projects you have marked as trusted. Project instructions use the AGENTS.md convention, and the /init command scaffolds one. MCP servers are configured in the same file and work across both the CLI and the IDE extension.
