Codex limits—5-hour and weekly usage limits, and six ways to continue after using them up

Codex limits are based on usage over a period of time, not a fixed number of messages. OpenAI’s published estimates for local messages per 5 hours with GPT-5.6 Sol are 10~100 for Plus, 50~500 for Pro 5x, and 200~2,000 for Pro 20x (OpenAI pricing documentation, checked September 23). A separate weekly limit may also apply. The same message uses more of the limit when the task is larger or the conversation is longer. As of September 2026, Plus has a 5-hour limit again. According to a post on X by Thibault Sottiaux, who leads Codex and ChatGPT engineering at OpenAI, Pro’s 5-hour limit is disabled for the time being, leaving only the weekly limit. The Pro information has not yet been reflected in the pricing documentation.

When you reach the limit, the current turn continues to completion within fair-use limits. Once it ends, the limit banner or usage page shows the options available for your account: add credits, apply a saved reset, upgrade, or wait for the limit to reset. Add two more options—switch to a lighter model as you approach the limit, and pay as you go with an API key, billed separately from your plan limit—and there are six in total. The table below puts all six options side by side and calculates the cost of using an API key on days you hit the limit, using actual rates.

Codex limits by plan — as of September 2026

This table reproduces the estimates in OpenAI’s pricing documentation and adds a column showing whether the 5-hour limit currently applies. The figures in the model columns are estimated ranges for local messages you can send in 5 hours. Only the “5-hour limit” column is based on sources other than the pricing documentation: the CLI changelog (Plus) and the X post (Pro), summarized below.

Because the limit is a range rather than a fixed number, OpenAI also says to check the current limit and reset time in the usage dashboard. The Korean Codex pricing page has the same table but does not specify the unit; “per 5 hours” is specified in the English pricing documentation.

Other things to know: local messages and cloud tasks share the same limit, and ChatGPT Work and ChatGPT for Excel also count against it. OpenAI only says that a weekly limit “may apply” and does not publish a number. Business standard seats have the same estimates as Plus, while figures for Free and Go have not been published. Prices are $20 per month for Plus, and $100 (5x) and $200 (20x) for the two Pro tiers (Pro tier help). However, since September 10, new sign-ups and upgrades to the Pro $200 plan have been temporarily paused, so the only plan you can newly upgrade from Plus to now is Pro $100 (existing Pro $200 subscriptions remain unchanged). Prices are shown in US dollars in OpenAI’s documentation; amounts on the checkout page may vary by region and taxes.

Is the “5-hour limit is gone” claim still true?

Only partly. Articles reporting the July news still appear near the top of search results, so here is the timeline in date order (dates are in US time).

• July 12 — Thibault Sottiaux, who leads Codex and ChatGPT engineering at OpenAI, posted on X that the 5-hour limit for Plus, Business, and Pro would be temporarily removed. Only the 5-hour limit was removed; the weekly limit remained in place (9to5Mac).
• August 25 — The 5-hour limit returned for Plus. The same person had announced this on X the previous day, US time, explaining that the change would balance compute demand and prevent Plus users from unknowingly using a whole week’s allowance. In the same post, they said that the 5-hour limit would remain disabled for Pro ($100 and $200) for “the next few months.”
• September 3 — The Codex CLI 0.153.0 changelog included an entry stating that “Plus and Team users get an earlier warning when their remaining usage for the approximately 5-hour window drops below half.” This is where the 5-hour limit for Plus can be confirmed in OpenAI’s documentation.

In short, Plus has both a 5-hour limit and a weekly limit, while Pro has only the weekly limit if Sottiaux’s post is accurate, because its 5-hour limit is disabled. However, the information about Pro appears only in the X post, and OpenAI’s pricing documentation still lists estimates for Pro per 5-hour period. The most accurate way to check what applies to your account now is to verify it directly as described below.

Check your remaining limit and reset time

# 터미널에서: CLI 버전과 로그인 방식 확인
codex --version
codex login status   # 지금 쓰는 인증 방식(ChatGPT 로그인 / API 키)을 출력

# 실행 중인 Codex CLI 세션 안에서: 남은 한도 확인
/status

# 계정 토큰 사용량 보기 · 저장된 재설정 쓰기 (ChatGPT 로그인 필요)
/usage

On the web, the Codex usage dashboard shows your remaining amount and reset time. In the ChatGPT app, these appear under Settings → Usage (in the Codex app, under Usage and billing). First, use codex login status to check how the CLI is signed in. If the account you used to open the dashboard is different from the one the CLI uses, the numbers will not match.

Reset times vary by account, and OpenAI Support cannot reset limits for you (OpenAI help). If you have a saved reset from a referral promotion or similar offer, applying it refreshes both the 5-hour and weekly limits and changes the weekly reset date. You can use it from the CLI with /usage, and unless the offer says otherwise, you must use it within 30 days of receiving it (referral promotion help).

When you hit a limit — six ways to continue

The current turn continues to completion within fair-use limits. Once it ends, the limit banner or usage page shows the options available for your account (wait, credits, a saved reset, or upgrade). Add switching models as you approach the limit and pay-as-you-go API key usage outside your plan, and the six options are as follows.

Credits are deducted at model-specific token rates, not dollar rates. For example, GPT-5.6 Sol costs 100 input and 500 output credits per 1M tokens (promotional rates in effect at least through November 21), but OpenAI only says that credit purchase amounts and prices vary by account, region, and plan. Credits expire 12 months after purchase and are not API credits (credit help). The Pro tier help says that upgrading to the Pro $100 plan applies the new limit immediately. If features that require a ChatGPT login, such as cloud tasks or GitHub code review, are blocked, an API key cannot unblock them, so choose credits, a reset, or an upgrade.

What does it cost per day to continue with an API key?

When you sign in with an OpenAI API key, Codex bills at standard API rates instead of using plan credits (OpenAI authentication documentation). If you set a compatible endpoint as the provider, billing uses that endpoint’s rates. Either way, there are no 5-hour or weekly limits; you pay for the tokens you use. Below are the costs for using a Kunavo key. Rates are read directly from the catalog, so the table updates when prices change.

“One task” is calculated as 20 steps for fixing one failing test. Each step uses 25,000 input tokens and 1,200 output tokens, so one task uses 500,000 input and 24,000 output tokens. The second line of the table header shows the input/output rates per 1M tokens. With GPT-5.6 Sol, one task costs $1.29, and a day of heavy use (20 tasks) costs $25.76. Using GPT-5.6 Terra for lighter fixes and switching to GPT-6 Astra only for bugs whose cause is unknown can substantially change the cost of the same day.

This table does not account for cache. Codex resends the conversation history with every step, so cached input is billed at 0.1x the input rate ($0.20 per 1M tokens for GPT-5.6 Sol), while the portion newly written to cache is billed at 1.25x. The more of the conversation history that hits the cache, the lower the cost compared with the table. If very little matches, most input may be billed as cache writes, raising the input cost to as much as 1.25x the table rate. Reasoning tokens for reasoning models are billed as output, so harder tasks use more output than the table assumes. Also, if a single request’s prompt exceeds 272K tokens, the entire request is billed at 2x input and 1.5x output rates, so don’t pack multiple tasks into one session; start a new one for each task. You can check the actual amount for each call in Usage.

So if you use Codex heavily every day, a flat-rate plan will generally cost less. The decision comes down to one division. The price difference for upgrading from Plus to Pro 5x ($100 per month) is $80 per month, which covers about GPT-5.6 Sol tasks at 62 per task (excluding cache). If you exceed your limit for fewer tasks than that in a month, it costs less to stay on Plus and use a key only on the days you go over. If you exceed it for more, Pro may be worth considering. The better the cache hit rate, the lower the cost per task and the higher the break-even number of tasks. The calculation comparing plan and API key rates is in Codex pricing.

Continue with a key only on days you hit the limit — two configuration lines

The Codex CLI and IDE extension work with both OpenAI API keys and Responses API-compatible endpoints. To use Kunavo, sign up, add at least $10 in credits, create a key on the API keys page, and add a provider block to your configuration file once. Leave the two lines commented out normally, and uncomment them only on days you hit the limit.

# ~/.codex/config.toml
# 평소: 아래 두 줄을 주석으로 둡니다 → ChatGPT 로그인(플랜 한도)으로 동작
# 한도에 걸린 날: 두 줄의 주석을 풉니다 → 다음 실행부터 Kunavo 키로 종량제
# 두 줄은 반드시 파일 맨 위, 첫 번째 [표]보다 앞에 둡니다
# model          = "gpt-5-6-sol"
# model_provider = "kunavo"

[model_providers.kunavo]
name     = "Kunavo"
base_url = "https://api.kunavo.com/v1"
env_key  = "KUNAVO_API_KEY"   # 키 자체가 아니라 「환경 변수 이름」
wire_api = "responses"        # 유일하게 유효한 값. 생략해도 같습니다

# env_key에 적은 이름의 변수에 키를 넣습니다 (키는 sk-kn-으로 시작)
export KUNAVO_API_KEY="sk-kn-..."

There are four common mistakes. In env_key, enter the name of the environment variable containing the key, not the key itself. Model names use hyphens as listed in the catalog (gpt-5-6-sol), so the dotted format used by your plan, gpt-5.6-sol, returns a 404. If a model line already exists, change its value instead of adding another one—duplicate keys cause a TOML syntax error. Also, the two lines must be at the very top of the file, before the first [표]. If you append the block to the end of a file that already has a table such as [projects."…"] for trusted projects, the two lines become part of that table and will not take effect. To switch back, comment out the two lines again. If model_provider is absent, Codex returns to the default openai and uses the login saved with codex login (usually a ChatGPT plan).

For installation through your first task, see How to use Codex. Each configuration setting is described in the Codex CLI integration documentation (English). To change the model for only the next run while using Kunavo, append -m as in codex -m gpt-6-astra.

When the Kunavo key path stops

• Balance — If your prepaid balance runs out, the request stops with 402 insufficient_quota. The balance does not expire (OpenAI credits expire after 12 months).
• Per-key monthly limit — You can set a monthly spending limit for each key on the keys page. If that month’s spending would exceed the limit, only that key stops with 402 and is not billed; other keys continue working (billing documentation (English)).
• Shared capacity — There is no dedicated quota or contractual SLA. If requests surge, you may get 429 errors. Don’t resend several requests at once; retry with gradually longer intervals (error documentation (English)).
• Failed requests are not billed.

The limits of the API key path are clear. Cloud tasks, GitHub code review, and Slack integration require a ChatGPT login and cannot be used with a key. Available models depend on the endpoint’s catalog.

Why limits run down quickly and how to use less

The cost of a message is not fixed. The model, context, reasoning effort, tool calls, and cache all affect usage, and each message gets heavier as a session grows longer. Fast mode uses the limit more quickly, and turns that include image generation use it 3–5 times faster on average. OpenAI recommends five approaches in its pricing documentation.

1. Make instructions precise and include only the necessary context. Remove irrelevant details.
2. Narrow down the materials it reads. Point it to relevant files and limit the scope.
3. Set the output length. Separate required work from optional improvements in your request.
4. Keep AGENTS.md concise. In large repositories, split it across subdirectories to reduce how much is injected at once.
5. Enable MCP servers only when you use them. Each server adds context to messages.

The Korean pricing page recommends one more approach: use lower-cost models for routine tasks. In the table above, GPT-6 Sol sends about three times as many messages as GPT-6 Astra in the same 5-hour Plus window. Check the dashboard every 1–2 weeks to learn your usage rate and anticipate when you might hit the limit.

Frequently asked questions

OpenAI’s documentation and X posts were checked on September 23, 2026. Limits and prices change frequently, so use your account’s usage dashboard as the source for current figures.