Guide

Codex rate limits, explained

Updated July 2026 5 min read

Codex comes bundled with ChatGPT plans, which makes its limits feel invisible — until an autonomous run stops halfway. Here's how the metering actually works, what burns through it fastest, and how to keep the numbers in sight.

How Codex is metered

Codex usage on ChatGPT plans runs on two meters at once: a rolling window of roughly five hours that starts counting when you start working, and a weekly cap that sits above it. The rolling window is the one you'll bump into on a busy afternoon; the weekly cap is the one that matters when you've had Codex grinding through tasks for days. Higher ChatGPT tiers come with substantially more of both.

The mental model

Two meters run in parallel, and the one closest to its cap is the one that stops you. Track which is tighter and when it resets — everything else is noise.

What drains the window fastest

Not all Codex work costs the same. Long autonomous runs — the kind where the agent iterates on a task for half an hour — draw continuously the whole time. Heavier reasoning modes multiply the cost of every step. And parallel sessions all draw from the same account budget, so three worktrees running at once burn the window three times faster than one. A quick question, by contrast, is nearly free. If your window vanishes "for no reason", it's almost always one of these three.

What happens at the cap

Codex pauses and tells you when the window resets. Your session survives, but anything the agent was doing stops mid-flight — which stings most when it happens deep into a run you kicked off and walked away from. The reset wait is usually tolerable if the rolling window stopped you; if the weekly cap did, waiting won't help today, and the real options are a higher tier or a second agent.

Reading the reset time

When Codex stops you, look at the reset countdown before deciding anything. Hours until reset means the rolling window is the binding one: schedule around it — grab lunch, review the diff it already produced, queue the next prompt for when it reopens. Days until reset means the weekly cap is what bit, and waiting is no longer a strategy: route the remaining work to an agent with a separate budget, or accept that your tier doesn't match your workload. Parallel worktrees deserve a special mention here — because every session draws from one account, a weekly cap hit often traces back to a week of three-at-a-time runs, not any single big task.

How to see limits coming

In the CLI, /usage opens the account usage view and /status shows the current session's token usage and remaining context — good snapshots, but you have to think to ask. The alternative is ambient: Vibe Island reads your Codex session and keeps both windows — percentage used and reset countdown — live in your Mac's notch. The step-by-step is in how to check your Codex usage.

5h 11% 4h1m 7d 2% 6h1m

A real crop of the Vibe Island notch, shown here with Claude Code — Codex gets the same treatment: both windows as live, color-coded numbers above your sessions.

Getting more out of your plan

Scope the runs. A tight, well-specified task finishes faster and burns less window than "figure it out" prompts that wander.

Save heavy reasoning for hard problems. Routine refactors don't need the most expensive mode.

Start big runs on a fresh window. Right after a reset, an autonomous task has the most room to finish without interruption.

Keep a second agent warm. Claude Code and other agents meter separately. Monitoring several agents at once turns "I'm rate-limited" into "I'll route this to whoever has headroom."

Common questions

How much Codex usage comes with a ChatGPT plan?

It depends on the tier — higher plans include substantially more Codex capacity. There's no fixed task count: heavy autonomous runs consume more of the window than short edits, so the percentage used is the number worth watching.

When do Codex limits reset?

The rolling window resets a few hours after you start drawing on it, on its own schedule; the weekly cap refills on a longer cycle. Run /usage in the Codex CLI for a snapshot, or keep a live countdown in view with a monitor.

Does Codex usage share a budget with regular ChatGPT?

Codex has its own metering on top of your plan. The practical takeaway is to watch the Codex windows specifically rather than assuming your chat usage tells you anything about agent headroom.

Can I track Codex and Claude Code limits together?

Yes — that's the point of a multi-agent monitor. Vibe Island shows each provider's windows side by side in your notch, so when Codex runs low you can see at a glance whether Claude Code has room to take over.