Repository navigation
Codex 5-hour usage limit is being consumed significantly faster since reintroduction #40880
Description
What version of Codex CLI is running?
OpenAI Codex (v0.149.1)
What subscription do you have?
Plus
Which model were you using?
gpt-5.6-sol medium
What platform is your computer?
WINDOWS
What terminal emulator and version are you using (if applicable)?
No response
Codex doctor report
What issue are you seeing?
Since the 5-hour Codex usage limit was reintroduced on Plus, I’m seeing a very significant increase in quota consumption compared with my previous usage of the exact same setup.
This is not just a perception caused by comparing the 5-hour limit with the weekly limit. I previously used Codex extensively with GPT-5.6 Sol while the 5-hour limit was active, and the consumption rate was clearly much lower.
With similar projects, prompts, reasoning settings and workflows, the new 5-hour allowance is now being consumed dramatically faster.
Could you please check whether there has been a change in quota accounting, compute weighting, token/tool-call weighting, or GPT-5.6 Sol consumption since the 5-hour limit was reintroduced?
I would also appreciate confirmation of whether this is expected behaviour or a potential regression.
What steps can reproduce the bug?
Uploaded thread: 01a03e46-cfd5-7f43-ab84-7b0a072380e8
What is the expected behavior?
No response
Additional information
I switched away from GPT-5.6 Sol to a lower-cost model to test whether Sol itself was responsible for the unusually fast depletion.
The 5-hour quota is still being consumed extremely quickly, including for very basic tasks such as small text/content modifications on a landing page.
This makes me suspect that the issue may not be model-specific, but could instead involve the effective size of the 5-hour allowance or its metering/accounting since the limit was reintroduced.
I had previously used Codex extensively under the old 5-hour limit, including on substantially heavier development tasks, and the practical amount of work available per 5-hour window was noticeably higher.
Additional controlled observation – GPT-5.6 Terra Medium
I now have a concrete example that suggests this is not specific to GPT-5.6 Sol.
The 5-hour allowance had just reset and showed 100% remaining before this task.
Model: GPT-5.6 Terra Medium
Task: basic landing-page content/link modifications, followed by targeted lint and production build verification.
Total working time: 6m 20s
5-hour allowance after completion: 84% remaining
In other words, this relatively basic task consumed 16% of the entire 5-hour allowance in 6 minutes and 20 seconds.
CLI output:
Worked for 6m 20s
gpt-5.6-terra medium · Context 68% left · 5h 84% left
The 5-hour meter was at 100% immediately before this request.
This reinforces my suspicion that the issue may involve the effective size or accounting of the 5-hour quota rather than GPT-5.6 Sol specifically.
Activity
github-actions commentedon Aug 26, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
RajashekarRaju commentedon Aug 26, 2026
+1
Damien-zeze commentedon Aug 26, 2026
16% of the 5-hour quota consumed in just 6 minutes with GPT-5.6 Terra Medium
Lucascristao commentedon Aug 27, 2026
O retorno do limite de 5 horas no Codex é muito restritivo. Antes, com o limite semanal, eu podia distribuir meu uso conforme minha necessidade e até concentrar grande parte da cota em um único dia. Agora posso ter quase toda a minha cota semanal disponível e ainda assim ser impedido de usar o Codex por várias horas. Seria muito melhor manter apenas o limite semanal, ou permitir que o usuário consuma livremente sua cota semanal.
PHAMQUOCVY commentedon Aug 28, 2026
I want to add a real-world data point from Codex Desktop on a ChatGPT Plus subscription.
This is not a complaint about Codex having limits. I understand that agentic coding consumes compute and that Plus cannot provide unlimited usage.
The problem is the effective amount of usable work that the current 5-hour window provides.
What happened
First 5-hour window
I started with a fresh allowance and used Codex for a real development task involving refactoring/fixing an eKYC implementation in my application.
After approximately 40 minutes of actual work, the entire 5-hour allowance was exhausted and Codex stopped my workflow.
After using the free reset
I then used the available free reset to start a fresh allowance.
After the reset, I performed only a limited amount of additional work, including a relatively simple document/Word export task.
That new 5-hour allowance has already reached approximately:
39% used / 61% remaining
Meanwhile, my weekly usage is only approximately:
6% used
So my sequence was essentially:
100% fresh 5-hour allowance → ~40 minutes of development → limit exhausted → free reset → limited additional work → another 39% consumed
I contacted OpenAI Support about this. Support explained that Codex usage is compute/token based and that large retained context, files, reasoning, tools, cached input, and speed/model choice can make later requests expensive.
I understand that explanation technically.
However, from a product and developer-experience perspective, the current behavior is still extremely disruptive.
A paid Plus user should not reasonably expect a feature described as a “5-hour usage window” to provide only around 40 minutes of practical development time under a normal agentic coding workflow.
The weekly limit is also nowhere near exhausted, yet the short-term window becomes the bottleneck.
The core UX problem
The issue is not that I expect unlimited compute.
The issue is that the current system makes sustained development unpredictable.
When I start a development session, I cannot tell whether I have:
several hours of usable Codex work;
one hour;
or only 30–40 minutes.
For an autonomous coding agent, this is particularly problematic because the limit can be reached while the agent is in the middle of reading files, modifying code, testing, debugging, and finishing a feature.
Please consider !
Codex itself is extremely useful. The problem is that the current rate-limit design is actively breaking the development workflow.
I’m sharing this because I want to continue using Codex — but an effective development session of approximately 40 minutes before being forced to stop for hours is not a sustainable experience on Plus.
SSefs commentedon Sep 1, 2026
I noticed an unusually sharp increase in Codex usage today on the Plus plan.
I ran one Ultra task that was moderate in scope — essentially an audit/review of a fairly large project with a significant amount of prompt/context data.
After a single working session:
• 5-hour quota: dropped from 100% to ~13% remaining (~87% consumed)
• Weekly quota: dropped by roughly 25%
What makes this unusual is the timing.
After the previous reset that addressed the usage-limit situation, my consumption became noticeably more stable and predictable. Similar workloads were consuming resources at a reasonable and consistent rate, and I did not observe anything close to this level of usage.
Today, however, I saw a very sharp change in consumption around midday. One audit task suddenly consumed most of my 5-hour allowance and approximately a quarter of my weekly quota.
Based on my previous usage patterns, the workload itself does not appear sufficient to explain such a large difference.
I’m not claiming that OpenAI changed the limits. This could potentially be a usage accounting/metering issue, especially given how abruptly the consumption behavior changed.
Has anyone else on Plus noticed a similar spike in 5-hour or weekly usage today?
LimpingNinja commentedon Sep 2, 2026
Not imagined, also happening here with Pro 20x quota. Down 20% in 2.5 hours when normally this light work I'm doing on bug fixes would have been a mere percent or 2 at most.
t-var-s commentedon Sep 2, 2026
I’m seeing a closely related issue on a different plan/platform, so adding this as another data point.
I’m on ChatGPT Business, using the Codex desktop app on macOS (Intel).
I reported the fuller reproduction in #20153, but the key sequence after my 5-hour limit reset was:
Fast mode was disabled.
The contrast is what caught my attention: the existing/complex thread plus compaction used 7% total, while a fresh thread after an app restart coincided with an 84-percentage-point drop.
So this does not appear to be limited to Windows/CLI or Plus accounts.
Full report/details: #20153
mharv1972 commentedon Sep 10, 2026
I’m seeing essentially the same regression.
I cannot provide a conventional A/B reproduction against the previous Codex state, because users obviously cannot rerun the same prompt against the historical service/model/metering configuration.
What I can provide is a strong historical baseline plus current controlled observations. I used GPT-5.6 Sol Medium extensively for comparable local repository work before this issue appeared, and exhausting the 5-hour allowance was historically very unusual for me.
Since first noticing the change on 7 Sept, materially similar work has repeatedly consumed very large portions of the allowance.
Some clean examples on GPT-5.6 Sol Medium:
52m09s run: 100% → 26% remaining
12m03s bounded compatibility fix: 18% → 6% remaining
bounded tablet UI corrective task: 100% → 70% remaining
subsequent small corrective tasks have consumed ~20–30% each
These are local WSL repository tasks, not Work/cloud tasks, with no external MCP servers and no relevant AGENTS.md.
The important regression for me is that Sol Medium itself now appears to consume the allowance dramatically faster than it did previously, while also taking much longer to complete comparable work.
I appreciate that this is not a laboratory A/B reproduction because the historical Codex backend is unavailable to users. The relevant comparison would need to be made by OpenAI from server-side telemetry against affected users’ historical sessions.
taobaibais commentedon Sep 12, 2026
Fresh Plus reproduction here on September 12, 2026 around 21:00 CEST.
My Codex 5-hour limit showed 100% remaining, I continued with only a few very small tasks/follow-ups, and shortly afterward the 5-hour meter showed 0% remaining / fully exhausted.
Crucially, I did not use GPT-6 Astra during this period. So the extreme depletion is still reproducible without Astra and appears consistent with the broader 5-hour metering/accounting regression described here.
Please check whether the effective 5-hour allowance, weighting, or delayed server-side reconciliation is still incorrect for some Plus accounts.
maxsdw commentedon Sep 22, 2026
I'm experimenting the same issue using gpt-5.6-sol high. I notice a significative increase of tokens consumptions since the 5-hour Codex usage limit was reintroduced for tasks of similar complexity.