Codex suits developers who want a multi-file change delegated end to end rather than typed line by line, replacing the job of reviewing AI-suggested diffs with reviewing finished pull requests. OpenAI's own materials put its failure rate on complex tasks at roughly 30%, so treat its output as a draft from a fast junior engineer, not a merge-ready patch.
What changed
OpenAI Codex is an autonomous coding agent that turns a plain-English task into a finished pull request: it works inside its own sandbox, edits files, runs the project's test suite and iterates on failures. Since September 22, 2026 it has run on OpenAI's GPT-6 models, now including GPT-6.1 Sol, behind a sandbox-and-approval safety layer rather than unrestricted machine access.
Maker: OpenAI · Autonomy: semi autonomous · Maturity: ENTERPRISE
Underlying models: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna
About OpenAI Codex
OpenAI relaunched Codex in April 2025 as an agentic coding system, retiring the branding from its 2021 code-completion API. Built by OpenAI, it takes a plain-English task, clones the target GitHub repository into an isolated sandbox, and works through the change itself instead of only suggesting code for a human to accept line by line. Adoption scaled fast enough afterward that OpenAI now treats Codex as a flagship agent product rather than a research preview.
Codex currently runs on GPT-6 Sol for complex, multi-step engineering work, GPT-6 Luna for smaller, high-volume edits and, since September 29, 2026, GPT-6.1 Sol, which OpenAI says delivers near-Astra intelligence at a fifth of the flagship's standard token prices (a vendor claim, not an independent test). All sit in the GPT-6 family OpenAI shipped on September 22, 2026 to replace the GPT-5.6 generation, below the flagship GPT-6 Astra; anyone weighing GPT-6 against rival model families can compare the options in Which LLM Should You Use in 2026. A sandbox layer enforces what files the agent can touch and which commands it can run at the operating system level, while a separate approval policy decides when it must pause for a human; an Auto-review mode introduced in 2026 lets a second reviewer agent clear most of those pauses automatically, which OpenAI says now accounts for the majority of Codex Desktop usage inside the company itself. Planning a multi-file change, running the test suite, iterating on failures, and opening the pull request without a person retyping each step is what separates Codex from an autocomplete-style coding assistant.
It fits software teams that want a change delegated rather than drafted: point Codex at a task and it will scaffold a new API or landing page, refactor a legacy module, generate unit tests, or diagnose a performance regression while a developer works on something else. That is a different shape of help than Cursor, an AI-native editor built for staying hands-on in the IDE rather than handing a task off. Support for Slack and a public-preview slot inside GitHub's Agent HQ let a non-coding teammate tag @Codex in a channel to ask a question or request a small fix, echoing what OpenAI's own ChatGPT Work targets for the same knowledge-worker audience, now roughly a fifth of Codex's user base. HokAI tracks it inside the coding-agent section of the directory alongside the picks in Best AI Coding Assistants in 2026.
There is no separate Codex subscription: it rides on the ChatGPT plan a person already has, from a $0 Free tier up through custom Enterprise pricing. Access comes through the ChatGPT desktop and web apps, a Visual Studio Code extension with roughly 9.8 million installs, an iOS app, and the open-source Codex CLI, which is Apache-2.0 licensed on GitHub.
OpenAI's DevDay on September 29, 2026 widened the product in four directions. Codex in the cloud now runs on a computer, remotely from a phone, or from any device, and reusable development environments let a team start tasks quickly from a shared setup with approved settings and permissions. The refreshed CLI starts and steers tasks by voice and adds an /agents view for tracking several delegated tasks at once. A new Code Review experience in the ChatGPT desktop app summarizes diffs and lets a person question Codex before commenting on GitHub pull requests or GitLab merge requests, and automatic reviews let Codex take a first pass in the cloud while the developer is away. Codex Security gained a cloud version that scans whole GitHub repositories on demand or on a schedule, checks new commits, removes duplicate findings and prepares fixes with the laptop closed. Separately, the Agents API now supports computer use and brings Codex's multi-agent capabilities into a developer's own application, covered in the Agents API pricing guide. OpenAI also introduced Ultrafast, a premium speed tier of up to 8x faster token generation (about 300 tokens per second) in Codex, available with GPT-6 Astra on the top Pro tier and Enterprise.
Codex is the wrong pick in a few cases. The fastest tier is reserved for the top Pro tier and Enterprise, so most paying subscribers cannot buy it. The computer-control feature works on macOS only, there is no self-hosted option, and every change still needs a human reviewer. A team that wants to stay hands-on inside an editor will do better with Cursor, and one that prefers another model family can compare Claude Code first.
Codex ships changes to its CLI multiple times a week: version 0.159.3 was the stable release on September 30, 2026, days after 0.157.0 added GPT-6 Sol and Luna to the model picker along with Amazon Bedrock support and 0.156.0 shipped a full-screen terminal UI and an always-on voice mode. That pace has a downside: in July 2026, users running Codex in full-access mode with sandbox protections turned off hit a bug where the agent's attempt to override its $HOME variable to build a temp directory instead deleted real files in their home folder, a defect several developers reported publicly before OpenAI shipped a fix on August 19, 2026 that verifies delete targets before running them. OpenAI has also pushed Codex beyond pure coding: a March 2026 plugin system and Codex Security, a dedicated agent for finding and proposing fixes to vulnerabilities, point toward the company's stated goal of making it useful well beyond developers. Rivals move just as fast: in late July 2026, Windsurf was folded into Cognition and rebranded Devin Desktop, a reminder that the coding-agent field Codex competes in keeps consolidating.
Pricing
Codex has no subscription of its own: it uses whichever ChatGPT plan a person already pays for, from Free ($0) and Go ($8) through Plus ($20), Pro at $100, $200 or $500 a month, Business ($20 per user a month billed annually, $25 billed monthly) and custom Enterprise and Edu pricing. GPT-6 Astra Ultrafast is limited to the $500 Pro plan and Enterprise. Usage beyond a plan's included quota draws down credits billed per token: on OpenAI's published Codex rates GPT-6 Sol costs 50 credits per million input tokens against 2.5 for GPT-6 Luna, so routing routine edits to Luna is the cheapest way to stretch an allowance. A separate OpenAI API key bills per token at standard API rates instead of through ChatGPT credits and skips the cloud features. See [ChatGPT](/hub/tools/chatgpt) for the plan table.
Key Features
- Autonomous PR delegation: Clones a GitHub repo into an isolated sandbox, writes the change across files, runs the project's own test suite, and opens a pull request without further input.
- Cloud tasks with reusable environments: Tasks can start on a laptop, be steered from a phone, or run unattended in the cloud, with environments a team defines once and reuses.
- OS-level sandbox with approval policies: A sandbox restricts what files and commands the agent can touch at the operating-system level, and a separate approval policy pauses it before anything outside that boundary.
- Auto-review mode: A reviewer agent clears most out-of-sandbox approval pauses automatically; OpenAI says this now accounts for the majority of internal Codex Desktop token usage.
- Code Review: Summaries and diff exploration in the ChatGPT desktop app, with an optional automatic first-pass review in the cloud for GitHub and GitLab changes.
- Codex Security Cloud: Repository-wide vulnerability scans that run when asked or on a timer, investigate each finding, merge duplicates and draft the fix.
- Ultrafast speed tier: Astra Ultrafast streams up to 8x faster (about 300 tokens per second) in Codex, limited to the top Pro tier and Enterprise, and draws included usage down 8x quicker.
- Voice-steered CLI with /agents view: Dictate a task into the terminal, then watch every delegated job in one /agents list.
- Slack and GitHub Agent HQ access: Teammates can tag @Codex in a Slack channel or thread, and GitHub added Codex to its Agent HQ system in public preview so it can pick up tasks inside GitHub itself.
- Codex SDK: A TypeScript SDK lets a team embed the same agent in its own tools with structured output parsing and session resumption, with more languages planned.
Strengths
- No separate subscription: Codex rides on the ChatGPT plan a person may already pay for, instead of billing on top of an existing AI subscription (see the cost FAQ for exact tiers).
- The Codex CLI's source is public on github.com/openai/codex, letting a team audit or fork the client instead of trusting a closed binary.
- The cloud sandbox runs multiple tasks in parallel in the background, so a developer can queue several changes and review pull requests as they land instead of babysitting one session at a time.
- The Visual Studio Code extension has roughly 9.8 million installs, giving Codex a distribution channel most standalone coding agents lack.
Weaknesses
- OpenAI's own materials acknowledge that roughly 30% of complex tasks still fail outright, so a developer still has to review, not just merge, what Codex produces.
- The feature that lets Codex operate a desktop computer directly works on macOS only, leaving Windows and Linux users without that specific capability.
- In July 2026, a bug let Codex delete real files in a user's home directory when running in full-access mode with the sandbox turned off; the fix took OpenAI about six weeks to land.
- There is no self-hosting option: every mode, CLI included, depends on OpenAI's cloud models, so a team cannot run Codex fully air-gapped.
- The speed tier worth paying for, Astra Ultrafast, is reserved for the top Pro plan and Enterprise, and it consumes included usage at 8x the Standard rate.
Frequently Asked Questions
What does OpenAI Codex cost in 2026?
Codex rides on a ChatGPT plan rather than billing separately: Free is $0 a month, Go is $8, Plus is the standard paid tier, and Pro comes in $100, $200 and $500 tiers. Business is priced per user, cheaper when billed annually, and Enterprise and Education pricing is negotiated directly. Usage past a plan's included quota is billed in credits.
Is there a free OpenAI Codex plan?
Yes: the Free and Go ChatGPT tiers include Codex with GPT-6 Luna at Standard speed in the desktop app, subject to rollout. Both cap usage well below Plus, which adds the web, CLI, IDE extension, iOS app, cloud integrations such as automatic code review and Slack, and GPT-6 Sol, so anyone doing more than small tasks will hit the wall and either wait or upgrade.
Which tools compete with OpenAI Codex?
The closest competitors are Claude Code from Anthropic, Jules from Google, Muse Code from Meta, and Cosine, a smaller independent vendor. Anthropic's agent is generally reported as more conservative on large multi-file refactors, Jules and Muse Code fit teams standardized on Google's or Meta's own stack, and Cosine suits a buyer who wants a focused coding vendor with no broader chat platform attached. Cursor is a different shape of product, an AI-native editor rather than a delegate-and-review agent, for teams that want to stay hands-on in the IDE.
Should I choose Codex or Claude Code?
Neither wins outright: Codex has the distribution edge, riding on a ChatGPT plan tens of millions of people already pay for, while Claude Code needs its own Anthropic subscription. Reviewers tend to find Codex quicker for greenfield work such as scaffolding a new API from nothing, while Claude Code earns praise for making fewer speculative edits on large, tangled refactors. Both ship as a terminal-first CLI with a cloud option, so the pick usually comes down to which model family a team trusts more.
How do I get started with OpenAI Codex?
Sign in with an existing ChatGPT account, since Codex uses that subscription rather than a separate signup. From there, install the Codex CLI from GitHub or add the Codex extension in VS Code, point it at a repository, and describe the task in plain English; a first run typically finishes inside the sandbox within a few minutes, after which you review the diff or pull request it opens.
Top Alternatives
- Claude Code: Pick Codex for fast greenfield scaffolding delegated end to end; pick Claude Code if your team is Anthropic-committed and wants an agent reviewers say plans large multi-file refactors more conservatively.
- Jules: Pick Codex if you're already inside the ChatGPT ecosystem and want Slack and GitHub Agent HQ access; pick Jules if your stack is Google's and you want a coding agent from that same vendor.
- Cosine: Pick Codex for breadth of surfaces (CLI, cloud, Slack, mobile, VS Code); pick Cosine for a smaller, coding-only vendor without a broader consumer chat platform attached.
- Muse Code: Pick Codex if you want a coding agent bundled into a ChatGPT plan; pick Muse Code if your organization already standardizes on Meta's model family.
HokAI guides covering OpenAI Codex
- ChatGPT Use Cases in 2026: What It's Actually For Now: ChatGPT gained Space, Pages, team tasks and a Meetings beta at DevDay 2026. Here is each new use case, the plan it needs, and where a specialist wins.
- Can ChatGPT Edit Videos? How to Edit Video with ChatGPT in 2026: Can ChatGPT edit videos? Not alone. See how the ChatCut and HeyGen plugins work, the install steps, plan limits and costs we verified in October 2026.
- ChatGPT Plans Compared (2026): Free, Go, Plus, Pro and Business: ChatGPT plans compared on 3 Oct 2026: Free, Go, Plus $20, three Pro plans at $100-$500 and Business seats. Limits, context windows and who should skip each.
- Meta Started Charging for Its AI Agents. The Framework Under Them Just Lost Its Creator to OpenAI: Meta's Business Agent stopped being free on Aug. 1, 2026. Meta AI hit 1.2 billion users, and the framework behind Meta's own agents lost its creator to OpenAI.
- Best Agentic AI in 2026: Pick the Job, Not the Generalist: Manus, Genspark and seven other agentic AI tools compared on GAIA scores, credit costs and price, plus when a narrow vertical agent beats a generalist.
- OpenAI Agents API Pricing: What You Pay, What It Replaces, and When Claude or Bedrock Fits Better: OpenAI Agents API pricing has no platform fee: you pay tokens, tools and sandbox time. Costs worked out, plus when Claude Managed Agents or Bedrock fits better.