Codex is included with every paid ChatGPT plan and with Free and Go at small volumes. As of August 2026 the ladder is Free $0, Go $8/mo, Plus $20/mo, Pro $200/mo, and Business at $20 per user per month on annual billing. There is no separate Codex subscription. OpenAI describes Pro as offering maximum Codex tasks. Usage is metered by a rolling five-hour window with weekly limits on top, and overflow is bought as credits.
- Codex is included with ChatGPT, from Free upward. There is no Codex-only plan.
- Plus at $20/mo is the real entry point; Pro is $200/mo and OpenAI describes it as offering maximum Codex tasks.
- Two meters run at once: a rolling five-hour window and a weekly cap.
- Past the included usage you buy credits, priced from a published per-token rate card.
- Model choice moves the meter more than anything else. Luna is 25x cheaper per token than Sol.
- The API is a separate rail with no windows, and it is what CI needs.
The price of every plan
Codex has no price of its own. You buy a ChatGPT plan and Codex comes with it, on the web, in the CLI, in the IDE extension, and on iOS. The plan sets how much of it you get.
ChatGPT plans and the Codex usage each includes, from the Codex pricing page, checked August 2026.
| Plan | Price | Codex included | Who it suits |
|---|---|---|---|
| Free | $0 | Quick coding tasks only | Finding out what it is |
| Go | $8/mo | Lightweight coding tasks | Occasional use, cheapest paid entry |
| Plus | $20/mo | Full access to the GPT-5.6 family | Individual daily agent work |
| Pro | $200/mo | Maximum Codex tasks | All-day agent use, parallel sessions |
| Business | $20/user/mo annual | Included, extendable with workspace credits | Teams wanting central billing and admin |
| Enterprise and Edu | Contact sales | No fixed rate limits, usage scales with credits | SSO, compliance, procurement |
| API key | Per token | Yes, with a key | CI, automation, products |
How the limits actually work
Subscription Codex is metered by two meters running at once, and confusing them is why people report hitting a limit they thought they had room under.
- A rolling five-hour window. Local CLI messages and cloud chats share it. It refills continuously as old activity ages out.
- A weekly limit on top. It resets on a schedule, not continuously, so a heavy Monday is still costing you on Thursday.
Plus messages per rolling five-hour window, by model. Published as ranges because cost scales with tokens, checked August 2026.
| Model | Plus, per 5 hours |
|---|---|
| GPT-5.6 Sol | 10 to 100 |
| GPT-5.6 Terra | 25 to 200 |
| GPT-5.6 Luna | 250 to 2,000 |
The three levers, in order of how much they move the meter: which model you picked, how much reasoning effort you asked for, and how much context each turn drags along. Fast mode is a fourth, and it consumes credits at a higher rate for the models that support it.
Credits: what you buy when the included usage runs out
When you exhaust the included usage, Codex does not stop selling you capacity, and it does not force an upgrade. Plus and Pro accounts buy credits; Business and Enterprise workspaces buy workspace credits. Your included usage is always spent first, then the balance draws down.
Codex credit rate card, credits per million tokens, checked August 2026.
| Model | Input | Output | Relative cost |
|---|---|---|---|
| GPT-5.6 Sol | 125 | 750 | The expensive one |
| GPT-5.6 Terra | 50 | 300 | 2.5x cheaper than Sol |
| GPT-5.6 Luna | 5 | 30 | 25x cheaper than Sol |
# mechanical edit with a complete spec
codex -m gpt-5.6-luna "rename Config.Timeout to Config.RequestTimeout across the repo"
# ordinary feature work
codex -m gpt-5.6-terra
# the thing that resisted two attempts
codex -m gpt-5.6-sol -c model_reasoning_effort="high"The API rail, and when it is the right answer
An API key is a completely separate billing rail. No windows, no weekly cap, no credits: you pay per token and the only ceiling is the spending limit you set in the console.
OpenAI API list price, US dollars per million tokens, short-context tier, checked August 2026.
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-5.6-sol | $5.00 | $0.50 | $30.00 |
| gpt-5.6-terra | $2.00 | $0.20 | $12.00 |
| gpt-5.6-luna | $0.20 | $0.02 | $1.20 |
| gpt-5.3-codex | $1.75 | $0.175 | $14.00 |
Use a key when
- CI and automation. A subscription covers interactive use by one person, not a build server.
- Anything you ship. A product needs a key with its own terms and its own budget.
- Very spiky usage. A quiet month costs nothing at all.
- A week where the weekly cap is gone. Keys have no windows, so a key is the overflow valve if you would rather not buy credits.
For a human at a keyboard the subscription wins decisively. A developer using an agent most of a working day generates token volume that costs several times $20 at list price, which is exactly why the flat fee exists.
Codex against Claude Code, on price alone
The comparable ladder, checked August 2026.
| Tier | Codex, via ChatGPT | Claude Code, via Claude |
|---|---|---|
| Entry | Plus, $20/mo | Pro, $20/mo |
| Middle | No separate published rung | Max 5x, from $100/mo |
| Heavy | Pro, $200/mo | Max 20x, from $200/mo |
| Team | $20/user/mo annual | $20/seat/mo annual |
| Metered | OpenAI API | Anthropic API |
Spending less without buying a smaller plan
| Lever | How | Typical effect |
|---|---|---|
| Model routing | -m gpt-5.6-luna for mechanical work | Largest single saving |
| Reasoning effort | -c model_reasoning_effort="low" | Large, and quality-neutral on easy tasks |
| Context hygiene | /new between unrelated tasks | Moderate, and it improves answers too |
A good AGENTS.md | Fewer wasted orientation turns | Moderate, and compounding |
| Fast mode off | Leave the priority service tier alone | Fast mode bills credits at a higher rate |
| Narrower prompts | Name the files; state what must not change | Fewer turns, smaller diffs |
Questions people ask
Nothing on its own. It is included with ChatGPT: Free and Go get small amounts, Plus is $20/mo, Pro is $200/mo with maximum Codex tasks, and Business is $20 per user per month on annual billing. Prices checked August 2026.
There is a free tier of it. The ChatGPT Free plan includes Codex for quick coding tasks, and Go at $8/mo includes it for lightweight tasks. Neither is enough for daily agent work, which starts being comfortable at Plus.
No. Plus at $20 includes expanded Codex usage. Pro is $200/mo and OpenAI describes it as offering maximum Codex tasks. Move up when you are hitting the weekly cap, not before.
Credits are the overflow rail. Your plan usage is spent first, and once it is gone, usage draws from a credit balance you buy. They are priced from a published per-token rate card: 125 credits per million input tokens on Sol against 5 on Luna, as of August 2026.
Almost always model plus reasoning effort plus context. A single high-effort turn on Sol against a large repository can cost what twenty low-effort turns on Luna against one file cost, and the published per-window message ranges are wide for exactly that reason.
Both are $20 at entry, but their heavy-use ladders differ: ChatGPT Pro is $200/mo, while Claude publishes Max tiers at $100 and $200. Choose on capability and fit rather than pretending the ladders match.
No. A subscription covers interactive use by one person. CI needs an OpenAI API key, and you should set a spending limit in the console before the first unattended run.
Yes. Fast mode is a priority service tier that trades credits for latency, and it consumes credits at a higher rate on the models that support it. Leave it off unless you are waiting on the answer.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.