Codex 5-Hour Limits Return 2026: What Plus Users Get

OpenAI has walked back the quiet throttle it applied to Codex earlier this month, and the OpenAI Codex 5-hour limit is back for ChatGPT Plus subscribers. For about a week, Plus users slammed into usage walls mid-refactor with no reset clock and no warning in the CLI — a change that was never announced, only discovered. The reversal restores the rolling five-hour window that made a $20/month seat viable for real agentic coding, and it lands alongside clearer messaging about how Work mode consumption is metered. If you build with Codex, this changes how you batch sessions, what you can ship in a day, and whether the jump to Pro still pays.

What’s new with the OpenAI Codex 5-hour limit

Codex usage for Plus accounts is once again metered on a rolling five-hour window rather than the tighter, effectively-daily budget that appeared without notice. Under the restored scheme, consumption is tracked continuously: once five hours have elapsed since a given batch of usage, that capacity comes back. There is no midnight cliff and no single daily allotment to ration. A developer who burns through their allowance at 9am is back in business by 2pm — the behavior that made the original model workable for people who code in bursts.

The throttle that triggered the backlash did two things at once. It compressed the ceiling on how many agentic turns a Plus seat could push, and it removed predictability. Developers reported hitting caps in the middle of multi-file edits, with the agent halting after it had modified some files but not others. That is worse than a hard, known limit. An agent that stops halfway through a refactor leaves your working tree in a state you must manually reconcile. The complaints that flooded the OpenAI developer forum and GitHub issues on the Codex CLI repo targeted determinism, not generosity.

Alongside the rollback, OpenAI clarified the two-tier structure that governs ChatGPT Plus Codex rate limits. The five-hour rolling window is the primary throttle, and a secondary weekly ceiling sits above it as a backstop against sustained heavy use. Most Plus users will only ever feel the five-hour window; the weekly limit catches outliers running Codex around the clock. Work mode — the newer agentic surface that lets Codex operate across connected tools and longer-horizon tasks — draws from the same pool. That detail explains why a coding budget evaporates after an afternoon of research tasks.

Why it matters

• Predictability beats generosity. A known five-hour reset lets you plan work in blocks. An unannounced cap does not, and it makes Codex unusable for anything time-sensitive — you cannot promise a Friday ship date against an unknown ceiling.
• The Plus tier stays viable for real work. At $20/month, Codex with a rolling window competes directly with Claude Code on a Pro plan and with Cursor’s mid-tier. Under the throttle, it did not, and many developers were pricing out the $200 Pro seat.
• Work mode and Codex share a budget. Agentic research or document tasks during the day spend the same allowance you need for evening coding. Treat them as one pool, not two features.
• Mid-task cutoffs are a correctness problem. An agent that stops mid-refactor leaves partial edits on disk. Structuring work into small, independently committable units is now defensive practice, not just hygiene.
• Session structure is a real optimization lever. Front-load planning into a single cheap turn and reserve the window for execution; you will ship measurably more per cycle.
• The reversal signals responsiveness — and volatility. OpenAI listened, but the episode confirms that Codex usage limits 2026 can change silently. Build workflows that degrade gracefully rather than assuming a fixed ceiling.

How to use the OpenAI Codex 5-hour limit today

1. Update the Codex CLI. Limit reporting and the reset-time header improved in recent builds; older versions may not surface your remaining window at all.

   npm install -g @openai/codex@latest
   codex --version

2. Check where you stand before starting a long task. The CLI reports your current window state, including when capacity returns.

   codex /status

   If the reset is less than an hour out, do planning and code reading now, and save heavy multi-file execution for after the window turns over.

3. Separate planning from execution. Planning turns are cheap; agentic execution turns are not. Get the plan out in one pass, review it yourself, then spend your window running it.

   codex "Read src/auth/ and produce a step-by-step migration plan
   to move session handling from cookies to JWT. Do NOT edit any
   files. Output the plan as a numbered checklist with the exact
   files touched at each step."

4. Make every step independently committable. This is the best defense against a mid-task cutoff. Tell the agent to commit as it goes so a stop leaves you at a clean boundary, not a half-edited tree.

   codex "Execute step 3 of the migration plan only. After the
   edits pass the test suite, stage and commit with a message
   prefixed 'refactor(auth):'. Then stop and wait."

5. Pin a scoped config so runs stay cheap and bounded. Narrow the agent’s file scope and cap its autonomy — fewer wasted turns sends more of your window to work that lands.

   # ~/.codex/config.toml
   model = "gpt-5-codex"
   approval_policy = "on-request"
   sandbox_mode = "workspace-write"
   
   [projects."/Users/you/dev/api"]
   trust_level = "trusted"

6. Add an AGENTS.md so you stop paying for re-explanation. Every session where the agent rediscovers your conventions is window spent on nothing.

   # AGENTS.md
   ## Commands
   - Test: `pnpm test`
   - Lint: `pnpm lint --fix`
   
   ## Rules
   - TypeScript strict. No `any`.
   - Never edit files in `generated/`.
   - Commit after each logical unit; never batch unrelated changes.

7. Batch around the clock. Note the timestamp when you first hit a wall — capacity returns roughly five hours later. Two disciplined windows per working day is a realistic Plus rhythm; three is achievable if you start early.

How it compares

Rate-limit structures differ more than raw pricing does, and the structure usually determines whether a tier works for you.

Plan	Price	Limit structure	Best for
ChatGPT Plus (Codex)	$20/mo	5-hour rolling window, plus a weekly backstop; shared with Work mode	Solo devs coding in bursts; a few focused sessions per day
ChatGPT Pro (Codex)	$200/mo	Same rolling structure, dramatically higher ceilings	All-day agentic work, parallel tasks, long-horizon runs
ChatGPT Business	$25–30/seat/mo	Per-seat allowance with pooled admin controls	Teams needing shared billing and policy control
Claude Code (Pro / Max)	$20 / $100–200/mo	5-hour rolling window plus weekly caps	Terminal-first workflows; long-context codebase reasoning
Cursor	$20–40/mo	Request/compute-based allowance, then usage pricing	IDE-native editing with tight inline feedback loops
Codex via API	Pay-per-token	No subscription cap; org rate limits only	CI pipelines, automation, unpredictable spikes

The ChatGPT Plus vs Pro Codex decision is simple. If you hit the wall more than once a day, every day, the arithmetic favors Pro. If you hit it a couple of times a week, restructuring your sessions beats a 10x price jump. And if your usage is bursty and automated rather than interactive, the API sidesteps subscription caps entirely.

What’s next

Watch whether OpenAI ships real-time limit visibility as a first-class feature. Most of the anger during the throttle week came from opacity, not scarcity — developers could not see what they had left or when it returned. A live remaining-budget indicator in the CLI and the ChatGPT interface would defuse most future incidents, and signs point to it being on the roadmap. Expect the Codex weekly limit reset to become visible too, since the weekly backstop is the least-documented part of the system.

The second thing to watch is how Work mode’s growth pressures the shared pool. As more agentic surfaces route through the same allowance — code, documents, connected tools, longer-horizon tasks — a single undifferentiated budget gets harder to reason about. The likely resolution is either separate meters per surface or a transparent credit system that shows what each action costs. Either would beat the current guesswork.

Expect continued churn. Inference capacity is finite and demand for agentic coding is growing faster than supply, so OpenAI Work mode limits and Codex ceilings will keep moving. Do not optimize hard against today’s exact numbers. Build habits that survive changes: small commits, explicit plans, scoped configs, and a fallback for when the window closes on you. Developers who structure work that way barely notice policy shifts. Everyone else re-learns the lesson each time.

Frequently Asked Questions

What exactly is the Codex 5-hour limit?

It is a rolling usage window rather than a daily quota. Codex tracks your consumption continuously, and capacity from a given period of use becomes available again about five hours later. There is no fixed daily reset time — the clock ties to when you actually used the service.

Did OpenAI increase limits, or just restore them?

This was a restoration, not an increase. OpenAI reversed a tightening it had applied quietly and returned Plus accounts to the previous rolling-window behavior. The headline number did not go up; the predictability came back, which is what developers were asking for.

Does Work mode usage count against my Codex limit?

Yes. Work mode and Codex draw from the same allowance on a Plus seat. Spending the morning on agentic research or document tasks reduces what is available for coding later. Plan them as a single budget.

How do I know how much I have left?

Run codex /status in the CLI, which reports your current window state and when capacity returns. Visibility is still coarser than most people want, so noting the timestamp when you first hit a wall is a useful manual backstop.

Is there also a weekly limit?

A weekly ceiling sits above the five-hour window as a backstop against sustained heavy use. Most Plus subscribers will never encounter it — the rolling window binds first. If you run Codex continuously across many days, the weekly cap is what will eventually stop you.

Should I upgrade to Pro?

Upgrade if you hit the wall multiple times a day, every day, and the interruptions cost you real working hours. If you hit it occasionally, restructuring your sessions — planning separately, committing per step, using an AGENTS.md — recovers most of the lost throughput at no additional cost. For automated or CI workloads, the API fits better than either subscription tier.