OpenAI · Pricing
Codex has no price of its own. It is included on every ChatGPT plan including the free one, expands at $20 a month, comes with a Business seat, and runs on a raw API key with no allowance at all. Past the included usage the unit stops being money and becomes a credit — and OpenAI publishes the rate card in credits while never saying what one costs.
Opens on a coding-agent day rather than a chat day — the inputs are set so the API column lands inside OpenAI's own published band for per-developer spend, printed further down this page. Change them to your volume; every plan re-judges as you type.
API side uses the vendor's published per-token list prices with prompt caching modeled (chat apps resend conversation history every turn — set cache to 0 for a stack without caching). Thinking models bill hidden reasoning tokens at the output rate — that's real spend in both worlds.
Plans cap usage in rolling windows and vendors don't publish those caps in tokens — verdicts compare the price of this volume, not entitlements. The higher the multiple, the more likely a cap cuts in first.
Most pages answer "what does Codex cost" with the price of ChatGPT Plus. That is the right answer for exactly one buyer and wrong for everyone else, because OpenAI sells access in four different units — per person per month, per seat, per token, and per credit — and gives it away in a fifth form that is not a trial. Every figure below is read off OpenAI's own pages on the date in the byline.
| Route | Bought as | Price | What it buys |
|---|---|---|---|
| ChatGPT Go | Subscription | $8/mo | Limited Codex access, the same wording the free plan carries, on GPT-6 Luna in the desktop app. No Codex Cloud, and credits only for the limited group of Free and Go accounts OpenAI has opened them to. |
| ChatGPT Plus | Subscription | $20/mo | First tier with expanded Codex usage, Codex Cloud, automatic code review and Slack, GPT-6.1 Sol, and credits to buy when the allowance runs out. |
| ChatGPT Pro 100 | Subscription | $100/mo | More usage than Plus — OpenAI no longer states a multiple — and, for now, no five-hour limit. Monthly billing only. |
| ChatGPT Pro 200 | Subscription | $200/mo | More usage than Pro 100. Subscriptions started after 29 September 2026 get a lower allowance than before; earlier subscribers keep the old one through 29 October. |
| ChatGPT Pro 500 | Subscription | $500/mo | The highest included usage, and the only personal tier with GPT-6 Astra Ultrafast — which spends the allowance at eight times the Standard rate. |
| ChatGPT Business seat | Per seat | $25/seat/mo $20/mo billed annually | Standard seats carry the same Codex usage estimates as Plus, then draw on a shared workspace credit pool if the workspace has bought credits. The usage-based Codex-only seat survives only in workspaces that added it before 24 June 2026. |
| Enterprise / Edu | Per seat | Contact sales | No published price. On flexible pricing there are no fixed rate limits — usage scales against a shared credit pool bought at the contract level. Some agreements bill usage in US dollars at token rates instead of credits. |
| OpenAI API key | API key | At model rates | Billed per token at API rates, with model availability following the key. Drops the cloud-side features: no GitHub code review, no Slack, no delegated cloud tasks. |
Codex is included on the free ChatGPT plan, described as limited access rather than a trial — it does not expire. The limits are not published as numbers, Codex Cloud starts at Plus, and buying credits is open only to a limited group of Free and Go accounts.
Read off, in order of appearance: learn.chatgpt.com/pricing, 2026-10-02 help.openai.com/about-chatgpt-pro-tiers, 2026-10-02 help.openai.com/using-codex-with-your-chatgpt-plan, 2026-10-02
The lightweight tier, described by OpenAI in the same words as the free one — limited Codex access rather than expanded. It buys headroom, not capability: the cloud-side features start one tier higher and credits are open only to a limited group of Go accounts, so this is a step for people who write small changes, not a cheap way to run an agent all day.
The first tier that is actually built for agentic work: expanded usage, the cloud surfaces, and the first where every account can buy credits when the five-hour window runs dry. It is also where OpenAI's own per-developer figure — printed further down this page, in the vendor's words — starts to look like the more relevant number than the subscription price.
Pro 100 lifts the Plus allowance — OpenAI no longer says by how much — and, for now, drops the five-hour limit altogether. Worth knowing what actually exhausts an allowance here: Fast mode uses it at 2.5 times the Standard rate, and Codex shares its pool with ChatGPT Work, ChatGPT for Word, Excel and PowerPoint and Workspace Agents, so heavy use of those eats the same allowance.
Pro 500 is the largest personal allowance OpenAI sells and the only personal tier with GPT-6 Astra Ultrafast — which uses the allowance at eight times the Standard rate, so the extra headroom goes quickly if you live in that mode. Pro 200 sits between the two, and anyone subscribing to it now gets a smaller allowance than its earlier subscribers. Before taking either, compare against the credit route: past a certain volume the question stops being which subscription and becomes how many credits a month, and the answer to that is priced in the section below rather than on the plan card.
Standard seats carry the same Codex usage estimates as Plus, plus what individuals do not get: a shared workspace credit pool with automatic recharge, usage alerts and usage reports for the owner. The usage-based Codex-only seat survives only in workspaces that added it before 24 June 2026, so for a team starting now the buying motion is standard seats plus credits.
No allowance to waste and no window to hit — and a genuinely different product, because the cloud-side features are not included on a key. Run your own volume before choosing it: for steady daily work the flat fee usually still wins.
The route that looks like the flexible option, and is also the one that quietly removes features.
On an API key Codex bills every token at the model's published rate — the rates live in our OpenAI rate table and move when OpenAI reprices. There is no plan allowance, no five-hour window and nothing to leave unspent, which is exactly right for work that arrives in bursts.
What you give up is not obvious from the pricing page. OpenAI's own description of the key route is that model availability follows the key and that it comes with no cloud-based features — GitHub code review, Slack and delegated cloud tasks are all plan-side. If your reason for wanting Codex is automatic pull-request review, the key route does not deliver it at any price.
The second thing that travels with this route is a different relationship to the credit system. Credits are a ChatGPT-account construct; on a key you are simply buying tokens, which is why the section below matters — it is the only way to compare the two sides in the same unit.
OpenAI meters Codex in credits and prices nothing in them. This is the only number on the page we worked out rather than transcribed, so the arithmetic is shown in full: the vendor's own rate card, divided by the vendor's own API price list, model by model.
| Model | API list, per 1M | Rate card, credits per 1M | Implied per credit |
|---|---|---|---|
| GPT-6 Astra | $10.00 in · $1.00 cached · $50.00 out | 250 · 25 · 1250 | $0.04 |
| GPT-6.1 Sol | $2.00 in · $0.10 cached · $10.00 out | 50 · 2.5 · 250 | $0.04 |
| GPT-6 Sol | $2.00 in · $0.20 cached · $10.00 out | 50 · 5 · 250 | $0.04 |
| GPT-6 Luna | $0.10 in · $0.01 cached · $0.50 out | 2.5 · 0.25 · 12.5 | $0.04 |
| GPT-5.6 Sol | $4.00 in · $0.40 cached · $20.00 out | 100 · 10 · 500 | $0.04 |
| GPT-5.6 Terra | $2.00 in · $0.20 cached · $12.00 out | 50 · 5 · 300 | $0.04 |
| GPT-5.6 Luna | $0.20 in · $0.02 cached · $1.20 out | 5 · 0.5 · 30 | $0.04 |
| GPT-5.5 | $5.00 in · $0.50 cached · $30.00 out | 125 · 12.5 · 750 | $0.04 |
| GPT-5.3-Codex | $1.75 in · $0.1750 cached · $14.00 out | 43.75 · 4.375 · 350 | $0.04 |
Rate card: OpenAI, 2026-10-02. API prices: our OpenAI rate table, which carries its own per-model verification dates.
The ratio is not approximately constant. It is the same on every model and every token type OpenAI publishes, which tells you the rate card was generated from the API price list rather than set separately. That has a practical consequence: a credit balance is an API budget in disguise, and any estimate you can make in tokens converts directly.
It also means the comparison people usually cannot make becomes straightforward. A subscription buys an allowance denominated in the same units your API bill would be, so the question "is the plan cheaper than the key" has a real answer for your volume rather than a directional guess — which is what the subscription-versus-API calculator is for.
Two cautions. First, this is what a credit buys, not what OpenAI charges you for one — the purchase price is not published, and we do not invent it. Second, the alignment is a fact about today's two tables: if OpenAI reprices one and not the other, the figure above stops being a single number and this section will say so rather than average them.
OpenAI publishes what a task costs in credits and never what a credit costs in money. Re-checked on 2 October 2026 — the Codex pricing page, the business rate card, the credits article for personal plans and the flexible-pricing article — and none carries a figure; the personal-plan article says purchase amounts and prices vary by account, region and plan.
These two figures are OpenAI's, not ours. They sit on the rate card, describe developers in aggregate rather than any particular one, and are the numbers the SaaS blogs in this search result quote without saying whose they are. We cite them because they are the only per-developer data anyone has published — and we attribute them, because adopting a vendor's arithmetic as a finding is how a price desk stops being useful.
The part vendors bury. Both of OpenAI's help articles on it were rewritten after our first read in August: the contradiction we reported between them is gone, and the refund terms are narrower than they were.
Cancel at least 24 hours before your next billing date to avoid the next charge; after an ordinary cancellation the paid features stay on until the end of the period you paid for. Cancelling refunds nothing by itself, uninstalling the app cancels nothing at all, and a subscription bought through Apple or Google Play is cancelled in that store rather than in ChatGPT.
Refunds are narrow. OpenAI's refund article says payments are generally non-refundable. A discretionary refund may be available on an unused individual subscription if you ask within seven days of the charge — with no substantial use, no repeat refunds and no open payment dispute — and there is no prorated refund after that. If a refund is granted, the access it paid for ends.
The regional rights sit on top. In the EU, the UK and Turkey a cancellation within 14 days of purchase carries a prorated refund; in the Republic of Korea a request within 7 days is refunded in full provided the service was not used, and a Korean refund request ends the subscription immediately.
Credits are the exception to all of it. They are non-refundable except for a verified billing error, unauthorised use or a confirmed OpenAI failure, expire twelve months after purchase, carry no cash value and cannot be transferred. Anything you top up is spent or lost, so top up for the month rather than for the year.
What people search next.