Last verified

OpenAI · Codex

OpenAI Codex.

OpenAI's coding agent · CLI, IDE, desktop, web and cloud · one ChatGPT account

Codex is OpenAI's coding agent: it reads a repository, edits files across it and runs commands from plain-language instructions. Unlike its closest rival it has no separate price and no paywall — every ChatGPT plan includes it, the free one included. What the free plan does not include is a published limit, and past your allowance the unit becomes a credit that OpenAI never prices in money.

Cheapest way in Free, then $8/mo
Free tier Yes
Ways to buy it 7
Oldest check 2026-08-19
§ 01 / WHAT IT IS

What it is.

Short version first: an agent that edits your repository and runs your commands, sold as a feature of a ChatGPT account rather than as a product with a price.

The thing to understand about Codex is that it is not one program. OpenAI ships it as a terminal CLI, an IDE extension, a mode inside the ChatGPT desktop app, a web surface, and a cloud runner that takes delegated tasks off your machine — plus code review on pull requests and a Slack entry point. They share one account and one usage allowance, and OpenAI's own documentation splits them into Codex Local and Codex Cloud, which is the split that matters for both permissions and price.

The CLI itself is open source under Apache-2.0 and installs with a single command, which reliably confuses the cost question. A free client is not a free tool: the licence covers the harness, and every turn still runs on OpenAI's models through either a ChatGPT plan or an API key. The two paths are not equivalent — on your own key the cloud-side features drop away, and code review, Slack and delegated cloud tasks are the parts you lose.

Which model runs underneath is configurable and not fixed. Workspace admins can set the starting model, the reasoning level and whether fast mode is available; individuals override per session where the workspace allows it. OpenAI has scheduled GPT-5.4 and GPT-5.4 mini to leave Codex for ChatGPT sign-ins on 31 August 2026, with GPT-5.6 Terra and GPT-5.6 Luna named as the replacements — your own API key is unaffected by that retirement.

One thing this page will not pretend. We have not run Codex ourselves: doing so means signing into an OpenAI account, and creating accounts is outside what this desk does. Everything below is read off OpenAI's own pricing, help-centre and repository pages on the date in the byline, plus two head-to-heads we worked through in full. Where a claim rests on documentation rather than a run, it says so — the same distinction our capability pages carry.

§ 02 / WHAT IT COSTS

What it costs.

Full pricing →

Seven ways to buy it and one way to get it for nothing, priced in four different units. The free tier is real and is not a trial; what it is not is documented as a number. The full breakdown, including what a credit turns out to be worth, is on the pricing page.

Route Bought as Price
ChatGPT Go Subscription $8/mo
ChatGPT Plus Subscription $20/mo
ChatGPT Pro (5×) Subscription $100/mo
ChatGPT Pro (20×) Subscription $200/mo
ChatGPT Business seat Per seat $25/seat/mo $20/mo billed annually
Enterprise / Edu Per seat Contact sales
OpenAI API key API key At model rates
§ 03 / WHEN NOT TO

When it is the wrong tool.

Five cases where the honest answer is something else. None of them is a complaint about quality — they are structural, they come off OpenAI's own pages, and no tier fixes any of them.

You need it to run without a network Every route in is an account with OpenAI — a ChatGPT plan or an API key — and both bill tokens processed on OpenAI's side. The CLI running locally is about where the harness executes, not where the model does. There is no offline mode and no local fallback, so a firewall that blocks the endpoint stops the tool rather than degrading it.
You want a bill you can predict Included usage is metered in a rolling five-hour window; past it, Plus and Pro buy credits and everyone else waits. Credits are pay-as-you-go, expire twelve months after purchase, are non-refundable except where the law says otherwise, and OpenAI's own terms state they carry no cash value and cannot be exchanged for money. That is a fine arrangement for uneven work and a poor one for a fixed budget.
Your code must stay out of model training on a personal plan On Plus and Pro, OpenAI's position is that conversations may be used to improve models unless you turn training off in data controls — an opt-out, not a default. Business, Enterprise and Edu are the other way round: inputs and outputs are not used to improve models by default. If the requirement is contractual rather than personal, the personal tiers are the wrong shelf entirely.
You are in the EEA, Switzerland or the UK and want every surface Two of the newer Codex features are geofenced. Sites is unavailable in the European Economic Area, Switzerland and the United Kingdom; Record & Replay excluded the EU, Switzerland and the UK at launch. Neither is core to editing code, but a European team evaluating Codex on a feature list will find part of that list does not apply to them.
You are a new Business workspace expecting a Codex seat Dedicated Codex seats closed on 24 June 2026 to new ChatGPT Business plans and to any workspace that had not already added one. Codex still arrives with a standard Business seat, but the separate usage-based seat type — and the buying motion built around it — is no longer available to teams starting now.
§ 04 / HOW IT COMPARES

How it compares.

All comparisons →

Two head-to-heads we have taken all the way through, both against tools people genuinely cross-shop. Each row says what the pair decides, because "which is better" is almost never the question anyone actually has.

Pair What it decides Verified
Codex vs Claude Code The like-for-like pair: two terminal agents from two frontier labs, so this one really is about model quality and price rather than shape. 2026-08-13
Codex vs Cursor The quality-per-dollar question. One agent scores higher on a neutral index; the other finishes faster and far cheaper per task, and the gap survives holding the model constant. 2026-08-13
§ 05 / ALTERNATIVES

Or something else.

See the list →

If one of the five cases above is yours, the list of what to use instead — open-source options and self-hosted ones included — is on its own page, with every competitor's entry price read off the vendor's site and dated.

Frequently asked.

What people search next.

Q · 01 Is Codex free? +
Yes, in a limited form that does not expire. OpenAI states that Codex is included across ChatGPT plans, including Free and Go, with usage limits varying by plan. What the free tier does not get is a published limit, the cloud-side features, or the ability to buy credits when the allowance runs out — those start at Plus.
Q · 02 How much does Codex cost? +
It has no price of its own. Access rides on a ChatGPT plan from $8 a month, on a Business seat, or on an API key billed per token. Our pricing page lays out all seven routes and works out what a credit is worth once the allowance is gone.
Q · 03 What is a Codex credit worth? +
OpenAI publishes the rate card in credits and never the price of a credit. Dividing that card by OpenAI's own published API rates gives the same figure for every model and every token type on it — the arithmetic, with both sources, is on the pricing page.
Q · 04 Is the Codex CLI open source? +
The client is. OpenAI publishes it under Apache-2.0 and it installs from npm, Homebrew or a shell script. The models it drives are not open, and the licence has no bearing on what running it costs — that is set by the plan or key you sign in with.
Q · 05 Which model does Codex use? +
There is no single answer: the default depends on your client version and configuration, and workspace admins can set a starting model, a reasoning level and whether fast mode is on. GPT-5.4 and GPT-5.4 mini are scheduled to leave Codex for ChatGPT sign-ins on 31 August 2026.
Q · 06 Does Codex work with my own API key? +
Yes, and it is a different product when you do. Billing moves to API rates with model availability following the key, and the cloud-side features — GitHub code review, Slack, delegated cloud tasks — are not included. OpenAI's own recommendation is to sign in with a ChatGPT plan.
Q · 07 Have you tested Codex yourselves? +
Not yet. Running it requires signing into an OpenAI account, which this desk does not do, so everything here is read off OpenAI's own pages and dated. Our tested work covers Claude Code, where 27 capability runs are published with transcripts.