Codex 5-Hour Limit: Meaning, Reset Window & Usage Estimates
The Codex 5-hour limit does not mean five hours of wall-clock coding time. It is a usage limit that applies across a five-hour window. Under official OpenAI documentation, a new five-hour window starts when you send your first message in Work or Codex after the previous window ends. OpenAI explicitly notes that users may reach this limit before five hours have passed depending on their tasks, model selection, reasoning settings, and context size. If reached during an active turn, Codex can continue working on that turn subject to fair-use limits before prompt submissions pause until the window ends.
Developers often assume the limit functions like a work timer that counts down only while typing or active in the terminal. When they reach a limit earlier during an active coding session, they assume the quota is malfunctioning.
OpenAI defines it as a five-hour usage window rather than wall-clock coding time. Because usage varies by task, model choice, settings, and context size, users may reach the limit before five hours have passed.
How a New 5-Hour Window Starts
Many developers wonder when the five-hour limit starts and resets. OpenAI provides the authoritative answer in Help Center Article 20001516:
“A new five-hour window starts when you send your first message in Work or Codex after the previous window ends.”
If your previous window has already ended, the next five-hour window does not begin until your next Work or Codex message. For example, if your previous window ended at 11:00 AM and you do not send your next prompt until 2:15 PM, your new five-hour window begins at 2:15 PM and will end at 7:15 PM. Once started, an active five-hour window continues running regardless of activity and does not pause during inactivity.
Why You Can Reach the 5-Hour Limit Before Five Hours Have Passed
OpenAI explicitly notes that users may reach the five-hour limit before five hours have passed. Usage varies depending on six documented factors:
Flagship models like GPT-6 Astra consume significantly more included usage per request than lightweight models like GPT-5.6 Luna.
OpenAI notes that reasoning effort can be adjusted independently. Higher reasoning settings instruct the model to perform more internal reasoning before responding, consuming more of your allowance.
Enabling Fast mode prioritizes speed and consumes your usage allowance more rapidly per task (see our guide on Codex Fast mode speed and multipliers).
Supplying larger inputs, such as entire repository files, long logs, or complex codebases, consumes more of your allowance on each request. In one controlled GPT-5.6 Sol test, roughly 6.5× more supplied input tokens coincided with 2pp vs. 7pp of observed 5-hour movement — see the full context-size comparison.
When Codex works through multi-step tasks that involve reading files, running terminal commands, and verifying code, each automated step contributes to your session's overall usage.
Generating extensive multi-file edits, comprehensive tests, or long explanations consumes more of your allowance than concise responses (see our diagnostic guide on why Codex usage drains fast).
First-Party Measurement: How Model Choice Moved the 5-Hour Meter
The factors above are OpenAI's documented explanation for why allowance depletes at different rates. To see what that looks like in practice, we ran a controlled experiment on ChatGPT Plus: the same frozen coding task, the same starting workspace, and the same prompt, run once each on GPT-5.6 Luna, GPT-5.6 Sol, and GPT-6 Astra at Medium reasoning effort in isolated Codex environments. Each run was graded by the same deterministic 112-test suite, with Codex Usage snapshots of the displayed 5-hour and weekly percentages taken immediately before and after.
| Model | Reasoning | Deterministic Result | Runtime | Measured 5-Hour Movement |
|---|---|---|---|---|
| GPT-5.6 Luna | Medium | 112 / 112 | 118,196 ms | No detectable movement (1-pt resolution) |
| GPT-5.6 Sol | Medium | 112 / 112 | 92,018 ms | 2 percentage points |
| GPT-6 Astra | Medium | 112 / 112 | 109,675 ms | 4 percentage points |
All three models passed every test in this run. Luna's run did not move the displayed 5-hour percentage at Codex Usage's 1-percentage-point resolution, Sol moved it by 2 points, and Astra by 4 — while Luna took the longest to finish. That ordering lines up with the model-selection factor above, but it is one controlled task, not a stable per-model multiplier: treat 0/2/4 as a single data point, not a formula you can apply to your own sessions.
This table reports Codex Usage's own observed measurement, not official OpenAI documentation. Each run used ChatGPT Plus, an identical frozen coding task and prompt, an identical starting workspace, and a clean isolated Codex environment, with one run per model and no retries. Results come from a single run per model and are not statistically significant; they should not be read as typical or guaranteed behavior for any model.
Published Local Message Estimates by Model (5-Hour Window)
OpenAI officially publishes estimated local-message ranges per 5-hour window on its Codex pricing page. These figures represent directional estimates based on typical coding sessions, not rigid prompt caps:
| Model | Primary Use Case | Plus ($20) | Pro $100 (5x) | Pro $200 (20x) |
|---|---|---|---|---|
| GPT-6 Astra | Deep reasoning, difficult debugging & architecture | 5–45 local messages | 25–225 local messages | 100–900 local messages |
| GPT-5.6 Sol | Feature implementation & balanced professional coding | 10–100 local messages | 50–500 local messages | 200–2,000 local messages |
| GPT-5.6 Terra | Everyday coding, docs & standard scripts | 25–200 local messages | 125–1,000 local messages | 500–4,000 local messages |
| GPT-5.6 Luna | Light edits, formatting & focused repetitive edits | 250–2,000 local messages | 1,250–10,000 local messages | 5,000–40,000 local messages |
| GPT-5.5 | Prior-generation general coding fallback | 15–80 local messages | 75–400 local messages | 300–1,600 local messages |
Message estimates vary widely because allowance is consumed based on task complexity rather than turn counts. Prompts with extensive code context, higher reasoning settings, or multi-step execution consume substantially more allowance per request than concise prompts with minimal context.
How the 5-Hour Window Interacts with the Weekly Window
OpenAI runs the 5-hour window alongside an overarching 7-day weekly limit. Understanding how they interact prevents unexpected downtime:
When your 5-hour window ends, your five-hour usage allowance is restored. However, the usage consumed remains accounted for in your weekly total. Your weekly allowance only resets on its scheduled 7-day anchor or via an applicable reset.
If your weekly allowance is exhausted, prompt turns pause even if your 5-hour window has available capacity. Conversely, if your weekly allowance has available capacity but you reach your 5-hour limit, prompt submissions pause until the five-hour window ends.
Plan Exceptions: The Business Premium 5-Hour Limit Removal
Not all paid plans enforce a 5-hour window. OpenAI documents a critical exception in Help Article 8792536:
The 5-hour limit is completely removed for Business Premium seats. Premium members have no five-hour limit and operate solely under a predictable weekly reset schedule with 5x more included usage than Business Standard seats.
Checking Your 5-Hour Status and Exhaustion Behavior
To view how much 5-hour headroom remains and when your current window resets:
/status in any session. Output includes:Frequently Asked Questions About the 5-Hour Limit
Does '5-hour limit' mean I get 5 hours of continuous coding time?
No. The 5-hour limit is not an elapsed wall-clock timer of coding time. It is a usage limit across a five-hour window. OpenAI explicitly notes that users may reach the limit before five hours have passed depending on their tasks, model selection, reasoning settings, and context size.
When exactly does a new 5-hour window start?
According to OpenAI Help Article 20001516, a new five-hour window starts when you send your first message in Work or Codex after the previous window ends. If your previous window has already ended, the next window does not begin until your next Work or Codex message.
Why do message estimates have wide ranges (e.g., 5 to 45 messages for Astra)?
OpenAI's published message ranges are directional estimates of typical usage per five-hour period, not fixed message caps. Usage varies based on the model chosen, reasoning settings, input and output sizes, and multi-step tasks. Tasks with large context and high reasoning consume significantly more of the allowance per request.
What happens when my 5-hour limit is reached?
If a usage limit is reached during an active Codex turn, OpenAI states that Codex can continue working on that turn subject to fair-use limits. Once that turn completes, subsequent prompt submissions pause until the five-hour window ends (as shown in /status or Settings > Usage). Your local files, Git commits, and session context remain completely intact.
Does using my 5-hour allowance also count against my weekly limit?
Yes. Model usage consumed during your five-hour window also counts toward your overarching 7-day weekly allowance. When the five-hour window ends, your five-hour allowance is restored, but your weekly allowance does not refill — it only recovers on your scheduled weekly reset date or through an applicable reset.
Does any plan remove the 5-hour limit completely?
Yes. ChatGPT Business Premium seats ($100/seat/month billed annually or $125 billed monthly) officially have no five-hour usage limit, operating with 5x more included usage than Business Standard under a predictable weekly reset schedule.
The rules and message estimates on this page are verified against official OpenAI documentation:
- Codex weekly limit guide — the 7-day cumulative usage ceiling and reset reanchoring rules.
- when does Codex usage reset — account reset timing and global usage refills.
- why is Codex usage draining so fast — troubleshooting rapid token consumption and runaway loops.
- Codex context window guide — session memory vs. subscription quota.
- ChatGPT Plus Codex limits — the $20/month tier in detail.
- ChatGPT Pro Codex limits — 5x ($100/mo) and 20x ($200/mo) tiers.
- ChatGPT Business Codex limits — Standard vs. Premium seats and the 5-hour limit removal.
- Codex usage limits table — master plan comparison matrix.