OpenAI Codex Usage Limits Explained | Generated by AI
Question: What does this Codex UI message mean — “剩余用量0%, 5小时, 1周, 0% 12:52AM 39% 7:10PM”, and why does it show two different time windows (5 hours and 1 week)?
Answer:
Your friend is hitting OpenAI Codex’s usage limits on a ChatGPT Plus subscription. Here’s what each part means:
The Two Rate Limit Windows
Codex (and ChatGPT Plus in general) enforces two independent rolling windows simultaneously:
| Label | Meaning |
|---|---|
| 5小时 (5 hours) | Short-term rate limit window — resets every ~5 hours |
| 1周 (1 week) | Long-term usage cap — resets weekly |
Both must be non-zero for you to use Codex. If either hits 0%, you’re blocked until the respective window resets.
Reading the Specific Values
剩余用量0% ← current remaining = 0%, fully exhausted
5小时 ← short window resets in 5 hours (at 12:52 AM)
1周 ← weekly window at 39% remaining (resets at some point in the week)
0% 12:52AM ← short-term bucket: 0% left, resets at 12:52 AM
39% 7:10PM ← weekly bucket: 39% left, last checked/updated at 7:10 PM
So the 5-hour window is the bottleneck right now — it’s at 0%. The weekly budget still has 39% left, but it doesn’t matter because the short window is empty.
Why Two Windows?
This is a standard token bucket / sliding window rate limiting pattern:
- Short window (5hr) — prevents burst abuse, e.g. running 1000 agentic Codex tasks back-to-back in an hour
- Long window (1 week) — caps total monthly-ish consumption per subscription tier
Codex is particularly expensive to run (it’s doing real code execution + LLM inference), so OpenAI applies tighter limits than regular ChatGPT.
“Codex 太烧token” (Codex burns tokens fast)
Correct. Codex CLI/agent mode is expensive because:
- Agentic loops — each task = multiple LLM calls (plan → tool use → observe → repeat)
- Long context — it reads your whole codebase/files into context
- Code execution feedback — stdout/stderr gets fed back into the next prompt
A single Codex session can easily burn 50k–200k tokens depending on task complexity. On Plus, the weekly allocation runs out fast if you use it seriously.
Bottom Line
Your friend needs to wait ~5 hours for the short-term window to reset. After that, they’ll have access again (with 39% of weekly budget remaining). If they hit 0% on the weekly window too, they’d need to wait until the weekly reset.
For heavy agentic use, Plus limits are too tight — Pro tier or API direct access (pay-per-token) is the practical path.