Why Is My Codex Usage Draining So Fast?
Unusually fast Codex usage can have several different causes. Some cases are normal high-cost workloads, while OpenAI has also acknowledged specific usage inefficiencies and its public Codex repository contains ongoing reports of background consumption and accounting discrepancies. The key is to determine which pattern matches your account before assuming your quota was reduced or that normal usage alone caused it.
If the exact same task burns 2x–3x more quota than last week, investigate potential platform regressions, cache hit variations, or accounting anomalies.
If quota drops when no prompts were sent, investigate desktop app background activity, automated extensions, or repository-reported passive drain.
If a single agent session drains a large chunk of quota, inspect context and tool-loop amplification across multiple continuations.
First rule out normal workload differences: changing the model or reasoning setting can materially change how quickly your allowance is consumed.
If CLI /status, web settings, and desktop dashboards disagree, a difference between displayed values does not by itself reveal whether the cause is delayed reporting, attribution, or accounting.
Official Service Status: As of September 9, 2026, the official OpenAI Status page reports no active incident affecting Codex usage limits or rate accounting. (Never assume an incident is active unless status.openai.com explicitly posts an ongoing advisory).
OpenAI-Acknowledged Inefficiencies: OpenAI staff in the developer community previously acknowledged specific issues that accelerated usage drain—including image handling during long sessions with multiple compactions, high p95+ Computer History consumption, conversation-title generation consuming more usage than intended, and degraded cache-hit rate periods—and stated that fixes had shipped for identified issues.
Open Repository Reports: In the official openai/codex GitHub repository, community issue reports regarding abnormal depletion, passive background drain, and accounting discrepancies remain open under meta tracker #41220. These represent user submissions and reproductions, not confirmed vendor defects.
Historical context: In June 2026, OpenAI investigated reports of Codex usage limits depleting faster than expected. OpenAI said some affected reports were related to abuse and fraud prevention systems incorrectly rate limiting certain accounts, while the observed impact appeared limited. The incident was later marked resolved and should not be assumed to explain current September symptoms.
Is Your Usage Drain Actually Abnormal?
Wall-clock time tells you almost nothing about quota consumption. Fifteen minutes of an autonomous agent running tool loops can process far more tokens than three hours of you typing.
To determine whether your drain rate is genuinely abnormal, establish a strict like-for-like baseline. Compare your current session against previous sessions that used:
A Plus plan ($20/mo) has an estimated 5–45 local messages per 5 hours on Astra, while Pro $100 provides 5x and Pro $200 provides 20x that capacity (see our limits table).
Changing the model or reasoning setting can materially change how quickly your allowance is consumed (compare presets in our recommended model guide).
Larger retained or selected context can increase input processing compared with a small task, but the exact context used on each turn depends on Codex's context-management behavior.
An interactive task with 3 manual turns differs fundamentally from an autonomous task that executed 40 automated file reads, bash commands, and test cycles.
Before concluding you are experiencing a quota defect, verify: Did you switch models? Did you enable Fast mode? Did an agent loop run 30+ tool iterations? Did you initiate tasks in ChatGPT Work or Excel that share your quota (see what counts toward Codex usage)? If all variables are unchanged and consumption remains drastically higher, explore the four diagnostic patterns below.
Usage Drops While Intentionally Idle
Losing quota while actively working is normal compute cost. Losing quota while your machine is idle or overnight is fundamentally different.
Several community reports in the official openai/codex GitHub repository describe passive depletion where allowance meters decreased without intentional user interaction. These reports remain under investigation and are not vendor-confirmed bugs:
A user submitted a controlled reproduction reporting that opening the ChatGPT desktop application and leaving it running without submitting prompts correlated with weekly allowance consumption. The report suggested background suggestion polling or initialization routines may have passively drawn compute.
A recent submission reported that weekly allowance and purchased credits continued to diminish while Codex sessions remained completely idle.
- Record exact starting 5-hour and weekly percentages before walking away from your computer.
- Ensure all CLI sessions, cloud tasks, and IDE extensions are closed.
- Observe whether percentages move over a quiet 1-hour or overnight window.
- Rule out other activity on your account that can share the relevant agentic allowance, such as Codex tasks, ChatGPT Work, ChatGPT for Excel, Workspace Agents, or activity from another signed-in device where applicable.
One Task Consumes an Extreme Percentage
Users report surprise when a single prompt or modest feature implementation burns 20% to 50% of their five-hour rolling quota.
In an active session this is usually context and tool-loop amplification: a large retained context, multiplied by many model and tool continuations, adds up to an enormous amount of processing.
Long-running agentic tasks can accumulate substantial model-processing volume across repeated continuations, especially when large context is retained or repeatedly reused. OpenAI states that larger codebases, long-running tasks, extended sessions, context, reasoning, speed, and tools can all increase usage. When an iterative workflow executes dozens of continuations while maintaining session state, cumulative input and output processing expands significantly over the lifetime of the task.
A detailed user-submitted reproduction documented a single-agent Terra Medium task running 76 execution turns. Across those turns, the session logged 10.1 million cumulative input tokens (approximately 98% cached). In that report, processing over 10M tokens across 76 turns coincided with consuming approximately 33% of the account's 5-hour allowance in under 18 minutes.
Understanding cached tokens: That report shows that very large volumes of mostly cached context can coincide with substantial quota depletion across many continuations. It does not establish how OpenAI weights cached input when calculating personal subscription percentages. OpenAI does not publish the exact weighting formula for cached tokens versus uncached tokens in subscription quotas. To manage model tier and reasoning overhead, review our recommended model and Power presets guide.
In a controlled GPT-5.6 Sol test, increasing supplied input context from 171,739 to 1,113,661 input tokens increased the displayed 5-hour movement from 2 points to 7 while both runs still passed 112/112 deterministic tests.
This is one controlled task on one model, not a universal token-to-quota conversion rate. See the full comparison table and methodology on our context window guide.
Drain Rate Suddenly Changes With Same Workflow
A bounded, abrupt shift in burn rate is diagnostically different from consistent high-cost usage.
When a developer works on the same codebase using consistent models, reasoning levels, and prompt patterns, a bounded before/after change in burn rate under otherwise similar conditions is worth investigating because it differs from consistently expensive work.
A user reported a bounded timeframe where their Pro weekly quota meter depleted roughly 2.4 times faster during routine development tasks before subsequently stabilizing back toward their prior observed baseline rate. This is a user-submitted report, not vendor proof of a platform defect.
In the official repository's cross-report meta tracker (#41220), numerous developers note similar abrupt rate changes. When evaluating a sudden rate change under consistent conditions, distinguish between OpenAI-acknowledged historical contributors (such as periods of degraded cache hit rates), user-reported accounting discrepancies, and unknown causes requiring investigation.
Displayed Numbers Do Not Reconcile
You check your meter, see a dramatic drop, but cannot find where the usage went in your chat history or token tallies.
Different usage surfaces can sometimes appear inconsistent in user reports. A disagreement between displayed values does not by itself reveal whether the cause is delayed reporting, attribution, recalculation, or accounting.
Chat-level stats only show the main conversation stream. Subagents, multi-step subroutines, and tool calls do not roll up cleanly into individual chat cards.
OpenAI does not publish the exact subscription weighting formula. You cannot reconstruct percentage drops purely from raw token counts.
For a deep dive into official dashboard views, 7D/30D token history, and chat reconciliation caveats, visit our Codex Usage Dashboard & Analytics guide.
OpenAI-Acknowledged Usage Inefficiencies
OpenAI staff in the developer community and official status updates have acknowledged specific technical contributors that caused unexpectedly rapid usage depletion:
OpenAI identified inefficiencies in sessions containing attached images that underwent repeated automated context compactions. In long sessions, attached images were unnecessarily re-encoded across multiple compaction cycles, inflating token burn. OpenAI subsequently announced fixes for this behavior.
OpenAI confirmed that tail-end (p95+) usage spikes were being caused by the Computer History feature, which logs desktop and browser interaction history into the agent's context.
An automated background routine responsible for generating descriptive conversation titles was discovered to be consuming more usage than intended. OpenAI deployed a fix to constrain this background process.
OpenAI noted periods where certain user cohorts experienced degraded prompt cache hit rates. When cache hit rates drop, the model must process full uncached context repeatedly, which OpenAI noted causes usage to deplete noticeably faster.
OpenAI staff reported in the developer community that targeted fixes have shipped for these specific identified mechanisms. Those fixes do not establish the cause of newer or unrelated depletion reports.
Normal Reasons Usage Legitimately Drains Quickly
Not all rapid quota loss is a bug. Several standard features are inherently compute-heavy:
/fast status.How to diagnose your own case
If you think something is actually wrong with your account, collect the evidence before you report it:
Run /status in the CLI or check chatgpt.com/codex/settings/usage. Record the exact 5-hour percentage, weekly percentage, and current local timestamp.
Document your current model (e.g., Terra, Sol, Astra), reasoning effort level, whether Fast mode is on, and which extensions, skills, or subagents are active.
Execute a single defined task (e.g., editing one specific file) or conduct an intentional 1-hour quiet period where no prompts are sent.
Re-check /status. Calculate the exact delta consumed. Distinguish active-task drain from idle loss.
Never share private source code, secrets, or confidential prompts in public GitHub issues or forums. Per OpenAI Help Center Article #11369540, if you believe your usage was counted incorrectly, contact OpenAI Support through the Help Center so they can investigate.
What to Do If It's Normal High Consumption
If your diagnosis reveals that your workload is simply heavy, practical steps can keep you coding without hitting limits:
Switch to a more efficient model configuration such as Terra Light or Sol Medium using our recommended model guide, verify your tier's baseline allowances in our limits table, and review how rolling and weekly cycles restore in our reset guide. If you need immediate capacity before your window restores, evaluate purchased credits or paid instant resets, or check the change log for recent policy updates.
Codex Usage Depletion FAQ
Direct answers to common questions about fast consumption, idle loss, and quota accounting:
Why is my Codex five-hour limit draining so fast?
Fast five-hour depletion can come from several different classes, including normal expensive workloads (such as frontier models or high reasoning), context and repeated agent activity across multiple continuations, shared usage across surfaces like ChatGPT Work or Cloud tasks, or reported accounting discrepancies and background activity. Diagnosing your burn rate requires comparing your active settings against baseline consumption to identify which category applies.
Can Codex usage disappear while I am not using it?
If you are not intentionally running Codex or another supported activity that shares the same agentic allowance, a falling meter during an idle period is worth investigating. User-submitted reports in the official openai/codex repository, including #37445 and #43223, describe passive weekly allowance or credit depletion while users reported no intentional Codex or Work activity. These reports are not confirmed by OpenAI as platform-wide defects, so first isolate other activity on your account before treating the movement as unexplained passive depletion.
Has OpenAI acknowledged fast Codex usage depletion?
Yes. OpenAI staff in the developer community have previously acknowledged specific usage inefficiencies, including attached image re-encoding during long compaction sessions, elevated p95+ Computer History consumption, a title-generation feature that consumed more quota than intended, and temporary cache-hit rate fluctuations that increased usage. OpenAI staff stated that fixes had shipped for identified issues. OpenAI also separately resolved a June 2026 status incident where abuse and fraud prevention filters prematurely rate-limited some accounts.
Can long contexts and repeated tool calls drain Codex usage quickly?
Yes. Long-running agentic tasks can accumulate substantial model-processing volume across repeated continuations, especially when large context is retained or repeatedly reused. OpenAI states that larger codebases, long-running tasks, extended sessions, context, reasoning, speed, and tools can all increase usage. When an iterative workflow executes dozens of continuations while maintaining session state, cumulative processing volume can deplete a significant portion of a five-hour allowance in a short period.
Can cached context still be involved in high usage?
Yes. User report #41369 shows that very large volumes of mostly cached context (10.1M cumulative input tokens across 76 turns, ~98% cached) can coincide with substantial quota depletion (~33% of a five-hour allowance in under 18 minutes). However, this observed correlation in a user report does not establish how OpenAI weights cached input when calculating personal subscription percentages, as OpenAI does not publish its internal quota weighting formulas.
Why did my Codex usage suddenly start draining faster with the same workflow?
When an identical codebase, model, reasoning level, and prompt pattern suddenly experiences accelerated quota burn, possible explanations include OpenAI-acknowledged historical contributors like degraded cache hit rates, user-reported accounting discrepancies, or unknown causes requiring investigation. Community meta tracker #41220 catalogs user reports of abrupt rate shifts without workflow modifications.
Can I calculate my Codex quota exactly from token logs?
No. OpenAI does not publish the exact token-to-percentage formula used to debit subscription five-hour or weekly limits. Local token logs can show raw input, cached-input, and output totals, but they cannot reconstruct the subscription meter. OpenAI says Codex usage varies with the model, where the task runs, task complexity, context, reasoning, speed, and tools.
What should I do if I think Codex counted my usage incorrectly?
Document your baseline: capture starting and ending 5-hour and weekly percentages via /status or your settings page, timestamps, the active model and reasoning settings, and whether any background apps were open. Per OpenAI Help Center Article #11369540, if you believe your usage was counted incorrectly, contact OpenAI Support through the Help Center so they can investigate.
Related Quota & Diagnostic Guides
Compare CLI /status, web settings, 7D/30D token history, and why chat usage does not always equal totals.
Choose the right model and reasoning effort to balance capability against quota burn rate.
Understand why Astra draws heavily from shared allowance and how Plus vs. Pro tiers differ.
See how ChatGPT Work, Excel, Cloud Tasks, and Voice share or separate from your limit.
How rolling 5-hour windows and weekly resets restore your allowance.
Distinguish normal subscription quota pauses from API 429 errors and spend ceiling limits.
Understand how 272k to 1M token memory, auto-compaction, and session accumulation differ from plan quota.