# Codex Usage Limits, Explained (Updated June 2026) How Codex limits work: rolling 5-hour windows, weekly caps on some plans, one shared usage pool, and on-demand credits. The full system, with links to OpenAI's live numbers. *Published 2026-06-12 ยท https://proxyllm.ai/blog/codex-usage-limits-explained* Codex ships inside every ChatGPT plan, Free through Enterprise, and OpenAI meters it with usage windows instead of a token bill. The system has four moving parts: a rolling window of roughly five hours that sets your short-term pace, a weekly cap on some plans that sets your ceiling, one shared pool that every Codex surface draws from, and on-demand credits that let Plus and Pro users keep going past a limit. The exact numbers vary by plan and model and change over time, so this page explains the machinery and points to developers.openai.com/codex/pricing for today's figures. ## The system at a glance | Part | What it does | Where it applies | | ----------------- | ------------------------------------------------------------------------ | -------------------------------- | | 5-hour window | Rolling short-term meter; sets how hard you can push right now | All plans | | Weekly cap | Second meter over a 7-day cycle; bounds sustained heavy use | Some plans, notably Plus and Pro | | Shared pool | CLI, IDE extension, cloud tasks, and code review draw from one allowance | Every Codex surface, per account | | On-demand credits | Paid, metered top-up to continue past a limit | Plus and Pro | Each part has its own deep-dive on this blog. This page is the map. ## The 5-hour window The short window is the limit you feel first. Usage is metered over a rolling period of roughly five hours: work you did early in the window stops counting against you as the window rolls past it, rather than everything resetting at a fixed hour. In practice you can push hard for a stretch, hit the ceiling, and watch capacity return on its own schedule. The window is a pace limiter, not a daily allowance. Spend it fast and you idle; spread the same work out and you may never see it. Mechanics, reset behavior, and observed patterns are in [Codex's 5-hour window](/blog/codex-5-hour-limit-explained). ## The weekly cap Some plans layer a second meter over a seven-day cycle. Interactive users rarely meet it. Agent users do, because parallel sessions and retries compound: there are consistent community reports of heavy agent work burning a weekly limit in a single day, which then means days of waiting unless another lane exists. The weekly cap is the constraint that turns limit management from an annoyance into an operations question. The full mechanics and the coping strategies are in [the Codex weekly limit](/blog/codex-weekly-limit-explained). ## What actually consumes usage Codex does not meter a simple count of prompts. It meters work, and that includes the model's reasoning time, tool calls, and retries. A one-line question costs little. An agent session that plans, edits files, runs the test suite, fails, and tries again costs a lot, even if you only typed one sentence. Reasoning time is usage: one deep task can draw more from a window than a dozen quick questions. Three things follow from that: - Higher reasoning effort settings spend windows faster on the same prompts. - Parallel sessions multiply burn, because each one plans and retries independently. - Every surface counts. CLI sessions, the IDE extension, Codex cloud tasks, and code review all pull from the same pool, and OpenAI counts Codex toward a shared agentic usage limit on your account. ## How limits differ by plan Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu, and Enterprise. The shape of the system is the same everywhere; the size of the windows is what scales with the plan. Free and Go are evaluation-scale. Plus carries a working developer with room to spare on interactive use. The Pro tiers exist for people who run agents. Business, Edu, and Enterprise allocate capacity per seat with admin controls on top. OpenAI publishes the per-plan rate details at developers.openai.com/codex/pricing, and we keep a plain-language comparison with workload fit in [Codex limits: Plus vs Pro vs Business](/blog/codex-limits-plus-vs-pro). ## On-demand credits When a Plus or Pro account hits a limit, OpenAI offers metered credits as a paid way to continue without waiting for the reset. Credits are bought in ChatGPT and drain per usage, which makes them a relief valve rather than a pricing plan. They are manual, open-ended, and, per community reports, faster to drain on agent work than most buyers expect. When credits make sense and when a second account beats them is covered in [Codex on-demand credits](/blog/codex-credits-explained). ## What happens at the limit Without preparation, a hit limit means one of two things: wait for the reset, or buy credits. Both are fine for a person at a keyboard and bad for anything automated. The operational answer is lanes. We built [Codex Hosted](/) around exactly this: when a connected account's window exhausts, requests fall back to a second ChatGPT account if you have one, then to your own OpenAI API key, and return to the subscription lane when the window resets. The request log names the lane that served each call, so overflow is visible the day it happens. The behavior end to end is in [what happens when you hit your Codex usage limit](/blog/what-happens-when-you-hit-codex-usage-limit). One boundary worth stating plainly: each connected account must be your own, with its own subscription. OpenAI's terms prohibit sharing an account between people, and OpenAI keeps final discretion over its services either way. ## Why this page has no message counts Every forum thread quoting "X messages per 5 hours" is a snapshot, and most are already wrong by the time you read them. OpenAI tunes Codex limits as models rotate and load shifts, and it says the numbers vary by plan and model. So we describe the system, which is stable, and leave the figures to the sources that update: developers.openai.com/codex/pricing and the help center at help.openai.com. For your own account, the CLI's `/status` command is the ground truth. Limits are the one part of plan-backed Codex that people fear before trying and stop noticing once they have a fallback. If you are sizing a plan against a real workload, the [calculator](/calculator) maps your current API bill to a tier in about thirty seconds.