Codex and Claude limits: when do they reset?

Codex says “limit reached,” Claude Code shows a reset time, and the browser seems to display something else. That can look like a contradiction even when every message is correct. The word “reset” is used for several separate boundaries. Once you classify the boundary, you can tell whether waiting is enough or whether a paid continuation is worth considering.

This article reflects official documentation checked on July 19, 2026. Plans, usage models, and promotions can change. Your account display therefore outweighs a time copied from a screenshot, a forum reply, or an older guide.

“Reset” can refer to four different boundaries

A five-hour window controls included usage over a short period. A weekly limit can cap consumption across several of those windows. The two can be reached independently: a renewed short window does not help if the weekly allowance has already been exhausted.

A context or length limit is different. It controls how much one conversation can process, not how much time remains in a subscription allowance. Starting a fresh conversation may help when the context is full, but it does not restore spent plan usage. Long coding sessions matter here because files, chat history, and tool results can increase both context and consumption.

API rate and spending limits form a fourth category. They belong to an API account and its billing controls. A ChatGPT or Claude subscription is not a general-purpose API balance. Mixing up these boundaries can lead to unnecessary waiting or accidental paid usage.

Codex: a five-hour window plus possible weekly limits

OpenAI's current Codex pricing page states that local messages and cloud chats share a five-hour window. Additional weekly limits may apply depending on the plan. That is why a universal claim such as “Codex always resets Monday at 00:00” is unreliable.

Consumption is not a fixed number of prompts either. Model choice, context, reasoning, tools, retrieval, caching, and task complexity affect how much one request uses. A small routine function may consume only a fraction of the allowance, while a long repository investigation with many tool calls can consume much more.

OpenAI also groups supported agentic features. Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents can draw from the same agentic usage and credit pool when those features are available to the plan. Other ChatGPT boundaries, such as voice or file-upload limits, are not automatically Codex limits. Read the label on the actual warning before choosing a remedy.

Claude Pro: session usage and a fixed weekly reset

Anthropic documents a session-based usage limit for Claude Pro that resets every five hours. Pro also has a weekly limit across models. Its day and time are assigned to the account and stay fixed regardless of when a new conversation begins or when the subscription started.

Claude usage is shared across surfaces. claude.ai, Claude Code, and Claude Desktop draw from the same subscription allowance. Moving from the terminal to the browser does not create fresh capacity, even if the two surfaces visualize consumption differently or one has no long conversation open.

Anthropic explicitly separates usage limits from length limits. If only one chat's context is full, a fresh conversation or context management can help. If the time-based or weekly allowance is spent, the reset shown by the account remains decisive.

Find the reset that actually applies to you

For Codex, open the usage dashboard. During an active CLI session, /status shows the current allowance. Read whether the notice refers to a five-hour window, a weekly boundary, credits, or another limit. Do not infer the reset from your last prompt because another Codex surface or supported agentic feature may already have used part of the shared window.

For Claude, open Settings > Usage. The page displays progress through the five-hour session, remaining time, weekly usage, and the next weekly reset. Compare it with the warning in Claude Code. If the displays appear inconsistent, first confirm that both surfaces use the same account and organization.

A useful diagnostic request follows the same discipline as your first good prompt: record the observed message, the affected surface, the account context, and the outcome you need. “Codex is broken” becomes a question that can be verified.

Credits, a banked reset, and API usage are different

OpenAI offers eligible Plus and Pro accounts the option to use credits after included usage is depleted. Credits are purchased additional usage. Auto top-up can buy more automatically, so enable it only with a deliberate target balance and spending cap.

A banked Codex rate-limit reset is not the same thing. It can appear as a promotion-dependent referral benefit, is not cash or API credit, and applies only to the eligible Codex boundary described by the offer. OpenAI's general terms say a banked reset normally has to be used within 30 days unless the specific offer says otherwise. The in-product offer is the source of truth.

Claude also offers usage credits for supported paid plans. Once included capacity is spent, further activity is billed at usage-based rates. Anthropic states that credits do not change the regular five-hour reset. They let you continue working at an additional cost; they do not reset the clock.

An API key is another billing path. It can support urgent local work, but it uses standard API pricing and needs separate keys, budgets, and security controls. It is not a free subscription workaround.

Choose the next action deliberately

OpenAI allows an active Codex turn to continue after the limit is reached, subject to fair-use limits, so work in progress can finish. Do not assume another turn will start afterward. Save a small, reviewable checkpoint early: tests, diff, open questions, and the exact next step.

When the task can wait, the displayed reset is the lowest-cost option. When short-window capacity is merely tight, a smaller model, tighter scope, or less irrelevant context can preserve what remains. A bounded agent task can also reduce unnecessary tool calls and open-ended exploration.

Credits make sense when the time saved justifies the price and a spending cap is in place. An upgrade is easier to justify for repeated saturation than for one unusual spike. Record the choice and its cost so an urgent exception does not silently become the default workflow.

Checklist for the next limit warning

• I recorded the exact warning, surface, account, organization, and plan.
• I separated five-hour, weekly, context, and API boundaries.
• I opened Codex Usage or Claude Settings > Usage.
• I checked which other surfaces draw from the same pool.
• I saved the diff, tests, and next step before the session ended.
• I compared waiting, a smaller model, credits, and API usage by urgency and cost.
• I set a spending cap and reviewed auto top-up before enabling paid usage.
• I treated a banked reset as a promotion benefit, not a balance.

Limits of any general answer

A message count from another account cannot predict your allowance. Similar-looking tasks can consume different amounts because context, models, and tools differ. A screenshot of a reset time proves only what that account displayed at that moment.

If a timer jumps or the product remains blocked after the stated time, check provider status, active organization, time zone, and current documentation. A support request is stronger after those checks and should include the exact message and timestamp. One community report cannot prove a platform-wide defect or its cause.

The reliable workflow is short: identify the boundary, read the official account display, understand the shared pool, and only then choose how to continue. That turns “When is the reset?” into a concrete decision about time, scope, and cost.