What we checked on OpenAI Codex
Writes code
Codex authors new code from a prompt across CLI, IDE, desktop app and cloud tasks, editing files in the repository it is pointed at.
Inspect code, make changes, run commands, and automate repeatable work without leaving your terminal.
Debugs and fixes
The docs give a bug-fixing workflow where Codex reproduces the failure locally, proposes a patch and runs checks, so it lands fixes rather than only describing them.
Use this when you have a failing behavior you can reproduce locally.
Writes tests
A documented workflow has Codex write a unit test for a selected function following the conventions of existing tests, in the IDE or the CLI.
Use this when you want to define the exact scope to test.
Reviews code
A dedicated reviewer reads a diff someone else wrote and reports prioritized findings without touching the tree; GitHub PR review is GA, GitLab merge requests are in preview.
Codex starts a dedicated reviewer that reads the selected diff and reports prioritized, actionable findings without changing your working tree.
Whole-codebase work
Cloud tasks take whole repositories and the local review pane spans several Git repositories, but no page states a repository size or file-count ceiling for a migration.
Choose repositories. Codex inspects the project, works through setup, and asks you for missing details or access.
learn.chatgpt.com/docs/cloudChecked
Ships it
Cloud tasks can commit or open a pull request, and the GitHub Action posts results in CI, but reviewing in chat explicitly does not merge and no page claims deployment.
Reviewing a pull request in chat doesn't post comments, approve it, or merge it.
Runs autonomously
Cloud tasks keep working unattended and the approval policy can be set to never stop, but the default local mode pauses at the sandbox boundary and asks.
For most work, start with Ask for approval. It lets ChatGPT work within the current workspace and pauses before reaching beyond that boundary.
Multi-agent
Codex spawns parallel subagents whose results it consolidates into one response; the docs warn each subagent does its own model and tool work, so runs cost more tokens.
ChatGPT Work and Codex can run subagent workflows by spawning specialized agents in parallel and then collecting their results in one response.
Agent permissions
A sandbox sets which files and networks commands may touch while an approval policy decides when it stops; modes run from read-only to danger-full-access with approvals off.
The sandbox defines which files and network resources ChatGPT can access. - Approvals determine when ChatGPT pauses before an action or sends the…
GPT
The model picker and --model flag select OpenAI GPT models, with GPT-6.1 Sol recommended for complex agentic coding and Luna for focused repeatable tasks.
For complex coding and agentic workflows, use GPT-6.1 Sol when available to your account and client. Use Luna for focused, repeatable tasks.
Model choice
You choose the model and reasoning effort per session, and can route requests through Amazon Bedrock or a company gateway, but every model the pages name is an OpenAI model.
Codex uses the configured model_provider to choose which models appear in the model picker
In your editor
A Codex IDE extension is a first-class surface, available on Plus, Pro, Business, Enterprise and with an API key, and shares MCP configuration with the CLI.
The ChatGPT desktop app, Codex CLI, and IDE extension support MCP servers and share MCP configuration for the same Codex host.
On the command line
The Codex CLI is the primary terminal surface, installed by script, npm or Homebrew, with an interactive TUI and a non-interactive codex exec mode.
Use Codex interactively or call codex exec from repeatable workflows and pipelines.
In your pipeline
An official GitHub Action runs codex exec in CI to apply patches or post reviews, and automatic cloud review can comment on a pull request before you open it.
Use the Codex GitHub Action (openai/codex-action@v1) to run Codex in CI/CD jobs, apply patches, or post reviews from a GitHub Actions workflow.
In a browser
Cloud tasks are started and reviewed from ChatGPT on the web, mobile or the desktop app, with each task getting its own workspace.
Give Codex coding tasks to work through in the cloud, from investigating bugs to building features. Review changes and continue from the web, mobile,…
learn.chatgpt.com/docs/cloudChecked
Self-hosted
The CLI, SDK and app server run on your own machine and can be pointed at your own gateway or AWS account, but the cloud execution environment is OpenAI-operated only.
Customers don't provide, host, or manage Work Cloud containers.
Open source
The CLI, both SDKs, the skills, plugins and app-server sources are on GitHub, while the IDE extension and Codex cloud are listed as not open source.
On the page: Open Source, 'Open-source components' table, last rows
Licence
Mixed: the CLI, SDKs and app server are developed in the open on GitHub, the IDE extension and cloud are not, and the repo refuses outside code contributions.
It doesn't accept external code contributions or pull requests.
Open-source runtime
The open-source CLI still needs credentials: a ChatGPT sign-in or an OpenAI API key, or AWS Bedrock credentials, which is the one path that needs no OpenAI account.
Users authenticate with a Bedrock API key or AWS IAM credentials. They do not use ChatGPT sign-in or OPENAI_API_KEY for this provider.
Runs in your cloud
Local surfaces can send model requests to OpenAI models in your own AWS account via Bedrock or your own LLM gateway, but cloud tasks and hosted Work are unavailable on that route.
Hosted ChatGPT Work on the web, Codex cloud, and features that depend on OpenAI-hosted cloud services or cloud-managed discovery aren't currently…
SDK languages
Two official SDKs: Python (3.10+, drives the local app server over JSON-RPC) and TypeScript (Node 18+, server-side), both controlling local Codex threads.
The Python SDK controls the local Codex app-server over JSON-RPC. It requires Python 3.10 or later.
Tracing and evals
Local execution emits OpenTelemetry events to a collector you configure, and Enterprise workspaces stream codex_log records; the docs warn neither is a complete record of every action.
Local execution also generates OpenTelemetry (OTel) events, which you can collect by configuring an OTel collector endpoint.
How it meters
Two models at once: a per-seat subscription that includes usage, plus credits metered in credits per million input, cached and output tokens once the included limits run out.
The rates below are for Standard speed, quoted in credits per million input tokens, cached input tokens, and output tokens.
Cost of one job
A developer-month starts at $8 (Go) or $20 (Plus); Plus buys roughly 15-160 local GPT-6.1 Sol messages per five hours, and the vendor says similar tasks can cost very different amounts.
Tasks that look similar can consume different amounts of your allowance. Model choice, context, reasoning, tool use, retrieval, and caching all…
Free tier
A standing $0 Free plan gives Codex on quick coding tasks in the desktop app with GPT-6 Luna, subject to rollout; no sales call is involved.
On the page: Pricing, Individual plan cards, Free card
API access
You can drive Codex from your own stack with the TypeScript and Python SDKs, codex exec, the app-server protocol and the GitHub Action, authenticating with an API key.
If you use Codex through Codex CLI, the IDE extension, or Codex cloud, you can also control it programmatically.
MCP server
Codex is an MCP client: it connects to STDIO and streamable-HTTP servers with OAuth. Exposing Codex itself as an MCP server for another agent was removed in favour of the app server.
The codex mcp-server command and standalone codex-mcp-server binary have been removed. Use the [Codex app server] for existing integrations.
Bring your own key
Sign in with your own OpenAI API key and pay API rates, or route requests through your own Bedrock account or company gateway credential instead of a ChatGPT subscription.
You can also sign in to the ChatGPT desktop app, Codex CLI, or IDE extension with an API key. Get your API key from the OpenAI dashboard.
learn.chatgpt.com/docs/authChecked
A company can buy it
Business is sold to organisations from two users up, and Enterprise/Edu is quote-only through sales with workspace-wide security controls.
On the page: Pricing, Business / Enterprise tab, Business card
Seat model
Priced per user per month with a two-seat floor: $20 per user billed annually, $25 per user when billed monthly, and no published seat ceiling.
*2+ users, billed annually. $25 per user per month when billed monthly.
Pooled budget
Business, Edu and Enterprise buy workspace credits that members draw on, and eligible Enterprise and Edu admins cap each user; individual Plus and Pro buy their own credits.
Business, Edu, and Enterprise plans with flexible pricing can purchase additional workspace credits to continue working.
Admin controls
Admins set Agent Security policies that constrain approvals, web search, MCP hooks and sandbox requirements, and can force the login method or lock users to one workspace.
In managed environments, admins may restrict how users are allowed to authenticate:
learn.chatgpt.com/docs/authChecked
Audit log
Enterprise and Edu workspaces stream append-only compliance records including codex_log events, but the docs say not every command, file operation or approval appears in an export.
Don't assume every shell command, browser interaction, app invocation, file operation, or approval appears in a customer-visible compliance export.
Single sign-on
SAML SSO comes with the Business plan alongside MFA; Enterprise adds SCIM provisioning and RBAC. The pages do not name OIDC for workspace sign-in.
A secure, dedicated workspace with essential admin controls, SAML SSO, and MFA
Opt out of training
Business, Enterprise and Edu workspace data is not used for training by default; the vendor makes that commitment for those plans and not for the individual Free, Go, Plus or Pro tiers.
Business, Enterprise, and Edu workspace data is encrypted in transit and at rest and is not used to train OpenAI models by default.
Data residency
Data and inference residency controls exist for Enterprise, but cover only eligible content, regions and workloads; residency also disables local computer access, and Work is unsupported with UAE inference residency.
Data residency and inference residency apply only to eligible content and supported workloads, subject to the organization's agreement, region, and…
Getting out
Member self-export is an admin-controlled permission and deleted chats are scheduled for deletion within 30 days, but retention differs per data class and credit expiry is not stated.
Deleted chats are generally scheduled for permanent deletion within 30 days, subject to published security, legal, and de-identification exceptions.
Certifications
The pages document encryption in transit and at rest, no training on business data by default and conditional HIPAA/BAA coverage, but name no SOC 2 or ISO 27001 certification.
HIPAA and Business Associate Agreement coverage depend on the features and agreement in use.
OpenAI Codex price and terms
Terms
- Free to start
- PartlyPartly
The terms in full
Standing $0 Free plan, not a trial, but limited to GPT-6 Luna at Standard speed in the desktop app and subject to rollout; image generation is not available on Free, and the plan feature matrix covers only Plus, Pro, Business, Enterprise and API key.
What needs a higher plan
- Business, Enterprise & Edu
- Pooled budget
- Enterprise & Edu
- Audit log, Data residency
For AI agents
- MCP server
- Partly
- API
- Yes
Where these prices come from
learn.chatgpt.com/docs/pricing, checked
OpenAI Codex alternatives
All 9 alternatives- This spot is openBe the first agent people see on OpenAI Codex's page.
- First on your rivals' pages
- Top of your category
- First in the home page list
- First in search results
- In comparisons and guides
- In the weekly newsletter
- Visits and AI reads each month
7 more alternatives
All 9 publish a price.
- JunieFree, then about $11 a seat a monthJetBrains coding agent for the terminal, JetBrains IDEs and CI, running on hosted or your own model keys.
- Muse CodeFrom $5 a monthMeta's terminal and CI coding agent, built on the Muse Spark model.
- OpenCodeFree, then $10 a monthOpen source AI coding agent for the terminal, desktop, IDE and browser, with any LLM provider.
- CursorFree, then $8 a monthCoding agent in a desktop editor, terminal CLI, cloud VMs and CI, priced per seat plus token usage.
Questions about OpenAI Codex
- How much does OpenAI Codex cost?
- OpenAI Codex costs $8 a month. From learn.chatgpt.com/docs/pricing, checked 24 October 2026.
- Does OpenAI Codex have an API or an MCP server?
- It has an API. It has a limited MCP server.
- Is OpenAI Codex open source, and can you host it yourself?
- Part of it is open source. You can host part of it yourself.
- What have we checked on OpenAI Codex?
- 38 facts, under these headings: What it does, how much it does unattended, which models it runs on, where you use it, whose machine it runs on, building with it, what it costs to run, buying it for a team, what happens to your code.
- How do we check OpenAI Codex?
- We hold 38 facts on OpenAI Codex, quoted from 17 pages. Facts and price last checked 24 October 2026.
Sources and changes
Last changed
Sources on learn.chatgpt.com17
- /docs/pricing
- /docs/enterprise/chatgpt-work-cloud-security
- /docs/amazon-bedrock
- /docs/codex-sdk
- /docs/auth
- /docs/cloud
- /docs/code-review
- /docs/codex/cli
- /docs/open-source
- /docs/permission-modes
- /docs/prompting
- /docs/agent-configuration/subagents
- /docs/enterprise/compliance-api
- /docs/enterprise/work-admin-faq
- /docs/extend/mcp
- /docs/github-action
- /docs/models
Is this yours?
AI assistants like ChatGPT and Claude opened The Agents Index pages 2,746 times in September. Claiming is free. You get this page's reader numbers and the verified owner mark. You confirm it with an email at openai.com.
Report an error if a fact is wrong.