Claude Code and Codex Usage Limits: Resets, Credits, and Practical Strategies
Running out of Claude Code or Codex usage in the middle of a refactor is frustrating, but the limit is easier to manage once you separate four concepts: included plan usage, reset windows, one-time banked resets, and paid credits. They are not interchangeable, and the options shown to one account may not appear on another.
This guide uses the current OpenAI and Anthropic documentation as its baseline. Exact allowances, eligible models, promotions, and prices can change, so your own usage dashboard is the final authority.
Start With the Usage Dashboard
Before trying to predict a reset, inspect what your account actually reports.
For Codex, use /status in an active CLI session or open Settings → Usage in a supported Codex client. OpenAI says the dashboard can show which allowance is exhausted, an available credit balance, and a reset time when one applies.
For Claude Code, use /usage to inspect plan limits or /status for session and authentication details. Anthropic also exposes usage information in account settings. The message displayed when a subscription limit is reached normally includes the relevant reset time.
Avoid relying on a screenshot from another user. Limits vary by plan, account, workspace, model, task type, and current product rules.
Why One Task Can Consume More Than Another
Neither service can be understood as a fixed number of identical prompts. Usage depends on the work being performed.
Common factors include:
- the selected model and reasoning effort;
- how much conversation history is carried into each turn;
- the number and size of files the agent reads;
- whether the task runs locally or in an isolated cloud environment;
- tool calls, code review, browser work, and other agent capabilities;
- the duration and complexity of the task.
A request to rename one variable is not equivalent to an agent reading an entire repository, planning a migration, changing twenty files, and running a full test suite.
Claude Code Usage Limits
Claude usage depends on how Claude Code is authenticated.
Subscription Usage
When Claude Code is connected to a Claude subscription, its usage is shared with other Claude surfaces covered by that plan. Work performed on Claude web, desktop, and Claude Code can therefore draw from the same allowance.
Anthropic documents session-based usage windows and additional weekly limits for subscription plans. Max plans provide higher capacity than Pro, but they are not unlimited. The next reset time shown in Settings → Usage is more reliable than a generic estimate because it reflects the rules currently applied to your account.
API Billing
When Claude Code uses an API key, usage is pay-as-you-go rather than deducted from a Pro or Max subscription allowance. The /cost command can display spend for the current API-backed session.
Check whether ANTHROPIC_API_KEY is set in your environment. Anthropic warns that this can cause Claude Code to use API billing even when you also have a subscription. Authentication determines the billing path.
Usage Limit vs Context Limit
A usage limit controls how much work your account can perform during a period. A context limit controls how much information one conversation can hold. Reaching one is not the same as reaching the other.
Use /compact when you need to continue the same task but the conversation has become large. Use /clear when moving to an unrelated task. Clearing between unrelated tasks prevents old files and messages from being resent as part of every new turn.
Codex Usage Limits
Codex is available through ChatGPT plans, with usage that varies by plan. OpenAI advises users to inspect the current pricing page and usage dashboard rather than rely on a permanent prompt-count table.
Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents can share an agentic allowance and credit pool when those features are available on the plan. A long Codex task can therefore affect the capacity available to another supported agentic surface.
When the included allowance is exhausted, your account may offer one or more of these choices:
- wait for the displayed automatic reset;
- apply an available banked reset;
- purchase or use credits;
- upgrade the plan.
The actual options depend on account eligibility, plan, workspace, and region.
What Is a Banked Codex Reset?
A banked reset is a one-time promotional benefit stored on an eligible account. It is not a cash balance, API credit, or permanent increase to the plan.
If a reset is available, OpenAI instructs users to open Settings → Usage, select the usage summary showing the reset, review its expiration, and confirm the action. A successful reset refreshes at least one eligible usage window. If there is nothing eligible to reset, it should remain available.
Important details include:
- eligibility and expiration depend on the specific promotion;
- future resets are not guaranteed;
- a reset cannot be transferred or exchanged for cash;
- the dashboard may take a short time to refresh after use;
- support does not provide courtesy replacement resets simply because one expired unused.
Do not press a reset merely because one is visible. Compare its expiration with your next automatic reset and your immediate workload.
Banked Resets and Paid Credits Are Different
Flexible usage credits extend supported agentic work after included usage has been consumed. They operate as pay-as-you-go capacity and draw from a credit balance.
A banked reset refreshes an eligible limit once. Credits pay for additional supported usage. Neither should be described as the other, and neither is the same as API credit unless the applicable offer explicitly says so.
When cost matters, check whether a client is about to use included allowance, purchased agentic credits, or a separate API account before starting a long task.
Monitor Claude and Codex With CodexBar
CodexBar is an open-source menu bar application that displays usage meters and reset countdowns for AI providers, including Claude and Codex. It is useful when you work across multiple tools and do not want to open every account dashboard repeatedly.
Current CodexBar features include:
- provider-specific session, weekly, and monthly usage displays where available;
- reset countdowns and optional quota warnings;
- multiple provider and account views;
- local cost-history scans for supported Claude and Codex workflows;
- status polling and incident indicators;
- a CLI for scripts and terminal output.
CodexBar is a third-party tool, not an official OpenAI or Anthropic client. Its documentation says local parsing is used by default and browser-cookie access is opt-in. Review its permissions and release notes before allowing access to account data or local logs.
Strategies That Preserve Useful Capacity
Separate Tasks Before They Grow
Do not carry an unrelated debugging history into a new feature. Start a fresh session or clear the old context. A focused conversation costs less and usually produces better decisions.
Match the Model to the Work
Use the expensive, deep-reasoning option when the task genuinely needs it. Routine edits, searches, and mechanical changes often work well with a lighter model. Reserve the strongest model for architecture, difficult debugging, and decisions where a wrong answer would be costly.
Point to Relevant Files
Give the agent a narrow entry point instead of asking it to read the entire repository. Mention the module, function, failing test, or error. Let the agent expand its search when evidence requires it.
Ask for a Plan Before a Large Diff
A short plan can prevent a large incorrect implementation. For work spanning several files, ask the agent to identify the files, assumptions, and verification steps before editing.
Use Worktrees for Parallel Tasks
Parallel sessions that edit the same checkout can overwrite or conflict with one another. Separate substantial tasks into Git worktrees and assign each session a clear ownership boundary.
Compare the Reset Time Before Using a Banked Reset
If the automatic reset is close and the promotional reset expires much later, waiting may preserve the banked reset for a more valuable coding block. If an urgent task is blocked for hours and the reset is eligible, using it may be reasonable. Base the choice on the dashboard, not a universal rule.
Stop During Service Incidents
If requests repeatedly fail or usage appears abnormal, check the official OpenAI or Anthropic status page. Do not assume that an outage guarantees compensation or a global reset. Official documentation says promotional and automatic resets depend on specific announcements and eligibility.
A Practical Preflight Checklist
Before starting a high-consumption task:
- Check
/status,/usage, or the account dashboard. - Confirm which account and billing method are active.
- Note the session and weekly reset times.
- Decide whether the strongest model is necessary.
- Start a clean conversation for an unrelated task.
- Point the agent at the smallest useful repository scope.
- Define the test or artifact that proves completion.
- Keep any banked reset for a moment when it creates more value than waiting.
Final Takeaway
The safest way to manage AI coding limits is to trust the account dashboard, keep context intentional, and distinguish included usage from promotional resets and purchased credits. Do not plan a project around rumors that a model launch or outage will refill every account.
Claude Code and Codex both provide tools for checking current usage. CodexBar can make those signals easier to see, but it should complement rather than replace the official account data.