Codex Usage Limits: 5-Hour and Weekly Caps

OpenAI launched Codex CLI as a terminal-native agentic coding tool in 2025, and the usage model it shipped with confused a lot of developers. The confusion comes from the overlap: you have a 5-hour window that resets automatically, and a weekly bucket that does not. Both run at the same time, and hitting either one blocks all new Codex requests until the relevant window clears. GitHub issues and the OpenAI developer forums have hundreds of threads from people who thought they were safe on the 5-hour clock and got blocked anyway by the weekly aggregate.

What is the Codex 5-hour usage window?

The 5-hour window is a rolling compute budget tied to your ChatGPT plan subscription. The clock starts on your first Codex request and resets exactly 5 hours later, regardless of how many requests you made in that window. Agentic tasks — where Codex plans, executes shell commands, writes files, and loops — are particularly expensive against this window because a single high-level instruction can trigger 30–50 tool calls, each carrying a full context payload.

Reading the 5-hour window with /status

$ codex /status   Codex CLI — usage status  ─────────────────────────────────────  5-hour window     ████████████░░░  78%  Resets in         1h 12m   Weekly consumption shown in /usage or  tracked externally (see npx whoburnedmore)

Notice that /status only shows the 5-hour window percentage. It does not surface the weekly aggregate in a way that lets you track your remaining weekly budget. That omission is widely reported as a usability pain point — you can see the short-term window but not the long-term one.

Agentic loops amplify your token burn faster than expected

When Codex runs in agentic mode, it re-sends the full conversation context on every iteration of its planning loop. A 50K-token context re-attached on 20 loop iterations costs 1M input tokens before a single line of output code is produced. This is the most common reason developers are shocked by how quickly they exhaust the 5-hour window.

How does the weekly Codex cap work?

The weekly cap is an aggregate limit that accumulates across all your 5-hour windows during a calendar week. OpenAI has not published exact token counts for each ChatGPT plan tier — the specific ceiling is intentionally opaque and has been adjusted over time. What is documented in support threads and GitHub issues is the pattern: on the ChatGPT Plus plan the weekly cap is relatively tight; on ChatGPT Pro it is substantially higher; on team or enterprise plans the limits are negotiated separately.

Why the weekly limit catches developers off-guard

The 5-hour window has a visible countdown, so you can feel it approaching. The weekly cap has no visible countdown — there is no built-in Codex command that tells you "you have 3M tokens left this week." You find out by getting blocked. This is exactly the gap that npx whoburnedmore fills: it reads your local Codex logs and shows you your weekly token total in absolute terms, so you can see the trend before it becomes a wall.

$ npx whoburnedmore  ↳ reading codex usage logs…   CODEX CLI — WEEKLY BREAKDOWN  mon   3.1M tokens    $9.40  tue   5.8M tokens   $17.60  wed   4.2M tokens   $12.70  thu   6.9M tokens   $20.90   ← limit hit  ─────────────────────────────────────  total  20.0M tokens   $60.60   fri–sun: no new requests (blocked by weekly cap)

The output above shows the exact pattern: three normal days, a big Thursday session, and then nothing for the rest of the week. Without per-day visibility into token counts, there is no way to diagnose this kind of pattern.

What happens when you hit a Codex limit?

When the 5-hour window is exhausted, any new Codex request — whether typed manually or triggered by an automated task — returns an error. The error message typically says your usage limits have been reached and asks you to wait. Agentic tasks that are mid-execution when the limit hits will abort and leave your working directory in a partial state, which can require manual cleanup.

1. 1

   Check /status immediately

   When you get a limit error, run codex /status right away. If it shows the 5-hour window at 100% with a countdown, you know exactly when to resume. If /status itself is blocked, that is a signal you may have hit the weekly cap.
2. 2

   Run whoburnedmore to see your weekly total

   Run npx whoburnedmore to get your per-day breakdown. If the weekly total looks high relative to your past weeks, you have almost certainly hit the weekly aggregate and need to wait for the weekly reset.
3. 3

   Recover mid-task agentic work

   If an agentic task aborted mid-execution, check your working tree with git status before resuming. Codex may have written partial files or staged incomplete changes. Clean those up before the next session to avoid feeding corrupted context into the next run.
4. 4

   Plan the next session around your remaining budget

   Use your whoburnedmore weekly totals to estimate how much headroom you have. If Monday through Wednesday already consumed 60% of what you burned in a typical full week, keep Thursday and Friday to lighter tasks like short Q&A, test generation, or documentation.

How does Codex compare to Claude Code on limits?

Both Codex and Claude Code use the same two-window structure — a short rolling window plus a weekly aggregate. The specific ceilings differ by plan and change over time, but the architectural pattern is identical. This makes cross-tool token tracking especially useful: if you run both, you can see which tool is consuming more of your overall weekly budget and route lighter tasks to the one with more headroom remaining.

For a detailed walkthrough of how Claude Code's equivalent limits work, see the Claude Code rate limits guide. For a guide to reading your full Codex history (not just the current window), see how to check Codex CLI usage.