ChatGPT Plans and Codex Access

How Codex access is tied to a ChatGPT plan, what that gets you, and where the authoritative numbers live.

This page deliberately quotes no prices or quota figures. Those change, and a stale number in a tutorial is worse than no number. The single source of truth is the official pricing page.

Codex is included with a ChatGPT plan

From the official Codex landing page:

ChatGPT Plus, Pro, Business, Edu, and Enterprise plans include Codex.

That is the important fact for the CLI: you do not buy Codex separately, and you do not need to provision an API key for normal use. You sign in with your ChatGPT account.

The pricing page also lists Free and Go as including some Codex capability. The Plus card is the one that explicitly names "Codex on the web, in the CLI, in the IDE extension, and on iOS." Surfaces and limits vary by plan — read that page rather than trusting a tutorial table.

codex login
codex login status    # exits 0 when saved credentials are present
codex doctor          # local diagnostic report

There is no codex status subcommand in the official CLI reference. For the account, model, and configuration of the current session, use /status inside the TUI.

Which plan, and why it matters

Rather than reproduce a table that will be out of date next month, here is what actually varies between plans, so you know what to look for when you read the pricing page:

What varies	Why you'd care
Included usage before limits apply	How long a heavy day of agent work lasts
Model access	Whether the newest or fastest models are available to you
Cloud execution capacity	How much you can push to codex cloud
Admin and governance controls	Whether you can enforce requirements.toml across a team
Seat management	Whether the plan is per-person or per-organization

Two concrete points that are documented and worth knowing:

• ChatGPT Pro includes access to GPT-5.3-Codex-Spark as a research preview.
• Business, Edu, and Enterprise are the plans where managed configuration matters — see requirements.toml for what an administrator can pin.

For anything numeric, read learn.chatgpt.com/docs/pricing.

Watching your own consumption

Codex reports usage in-session, which is more useful than checking a dashboard after the fact.

/usage              # current session
/usage daily
/usage weekly
/usage cumulative

/status shows the session's account, model, and configuration state.

Two habits keep consumption predictable:

Match reasoning effort to the task. model_reasoning_effort runs from minimal through xhigh. High effort on a one-line rename is waste; low effort on a subtle concurrency bug is false economy.

model_reasoning_effort = "medium"

Don't carry dead context. A long session that drifted off-task spends budget re-reading history that no longer matters. /clear between unrelated tasks is the cheapest optimization available.

Where the API is separate

Two things are billed through the OpenAI API rather than your ChatGPT plan, and both are opt-in:

• A custom model provider. If you configure model_providers to point at your own endpoint, that endpoint's billing applies.
• High-volume image generation. Image generation normally draws on your included limits — the documentation notes it consumes them 3 to 5 times faster than text. Setting OPENAI_API_KEY switches large batches to API pricing instead.

Under normal CLI use with codex login, neither applies. If you have never set OPENAI_API_KEY and never touched model_providers, your usage is on your ChatGPT plan.

There is no api_key setting in the Codex configuration reference. Authentication is codex login, or a provider entry with env_key naming an environment variable. Older guides sometimes show an api_key key in config.toml; it is not present in the current official reference.

Enterprise and team deployments

If you are rolling Codex out to a team, the relevant surface is not the plan page but the managed-policy layer:

Concern	Mechanism
Restrict which approval policies are selectable	allowed_approval_policies in requirements.toml
Restrict sandbox modes	allowed_sandbox_modes
Restrict web search	allowed_web_search_modes (disabled is always allowed)
Allowlist MCP servers	mcp_servers entries with an identity block
Pin feature flags	[features] in requirements.toml
Data residency	enforce_residency

Managed permission-profile allowlists require Codex 0.138.0 or later. Clients on 0.137.0 and earlier silently ignore allowed_permission_profiles and managed default_permissions — a rollout that assumes enforcement without checking client versions is not enforcing anything.

See the Configuration Reference for the full requirements.toml surface.

ChatGPT Chat (conversation)

The Claude.ai counterpart is Chat, not a separate install. Official docs split the same app into Chat / Work / Codex (Use ChatGPT).

Choose	When	This guide
Chat	Questions, brainstorms, short drafts, settling a design	This page
Work	A file someone can open and review	ChatGPT Work
Codex	Repo, diff, tests, PR	Product line · CLI

Chat surfaces frontend engineers actually hit:

• Projects — chats, files, and instructions under one topic.
• Voice — desktop and iOS; can talk through uploaded files and Projects.
• Library — reuse saved files without uploading again.
• Web / desktop / mobile — one account. On 2026-07-09 the standalone Codex app merged into the ChatGPT desktop app. Every plan, including Free, can open Chat, Work, and Codex; limits still come from the pricing page.

The GPT-5.6 Sol slider in Chat does not change model behavior in Work or Codex.

Using ChatGPT and Codex together

They are different tools with different strengths, and the pairing is genuinely useful:

Task	Better in
Thinking through an approach before any code exists	Chat
Comparing libraries or architectures	Chat
An eight-slide deck / comparison sheet / recurring agenda	Work
Reading and changing files in a real repository	Codex
Running tests and iterating until green	Codex
Explaining a diff to a colleague	Chat
Automating a task in CI	Codex (codex exec)

A workflow that works: settle the design in Chat, write it down as an AGENTS.md entry or a task description, then hand the implementation to Codex with a verification command. The design conversation and the implementation session have very different context needs, and keeping them separate produces better output from both.

Atlas: officially retired — do not document it as current

The standalone ChatGPT Atlas browser is not a rumor to hedge. Official pages already closed it:

• Evolving Atlas into ChatGPT: browser-based agentic work moves into ChatGPT and Codex; Atlas is scheduled to stop working on 2026-08-09.
• ChatGPT is now a partner for your most ambitious work (2026-07-09): the Codex app merges into the new ChatGPT desktop app; “We'll begin sunsetting the standalone Atlas browser.”
• ChatGPT Release Notes, Retiring Atlas, repeats the same stop date.

Today (2026-08-18) is after that date. For browser-agent work use the desktop built-in browser, the Chrome extension, or Work’s cloud browser. Do not cite WSJ / Reddit merger rumors; those three pages are the record.

Whether Atlas bookmarks / passwords can still be exported from a leftover local install after 2026-08-09 is whatever the help article currently says. This guide does not invent steps.

Related pages

• Learning Map — family tree and decision tree
• ChatGPT Work — reviewable deliverables, plugins, Sites
• Codex CLI — installation and core features
• Codex Product Line — CLI, IDE, app, and cloud
• Codex Cookbook — task recipes
• Codex Cheatsheet — commands and config keys

Official sources

• Pricing — the only authority on plans and quotas
• Use ChatGPT — Chat / Work / Codex
• Models — model list and reasoning effort
• Configuration Reference — requirements.toml and every config key
• Evolving Atlas