OpenAI Codex users are raising concerns about how quickly their weekly usage allowances are being consumed, with several developers reporting that their available capacity appears to disappear faster than expected.
The reports do not establish that OpenAI has secretly cut Codex limits in half. However, multiple complaints on the official OpenAI Codex GitHub repository describe unusually rapid quota consumption, inconsistent usage readings and difficulties understanding how individual coding sessions affect weekly allowances.
One of the most detailed reports came from a ChatGPT Pro user who said their weekly allowance reset to 100% on August 1. The user reported that it fell to 91% during a roughly 4.5-hour GPT-5.6 Sol engineering session involving two bounded subagents. They also noticed a difference between local telemetry and the account dashboard.
Another Pro user reported that 75.25 million tokens consumed about 26% of their weekly allowance shortly after the same August 1 reset. According to the report, about 96% of those tokens were cached reads, leading the user to estimate that their effective capacity was substantially lower than before.
These reports are individual user experiences and should not be interpreted as proof that OpenAI deliberately reduced Codex quotas. They do, however, show that some developers are experiencing behavior they find difficult to reconcile with their expectations.
Developers Want Clearer Codex Usage Information
The central issue for many developers is not simply the size of their allowance but how usage is calculated.
OpenAI has moved Codex to token-based pricing. Its current rate card explains that consumption is calculated using input, cached input and output tokens rather than a simple per-message system.
This means that two coding requests that appear similar to a user can potentially have very different resource consumption. Large contexts, complex tasks, tool calls, reasoning and additional agents can all contribute to higher usage.
OpenAI also provides usage information through its Codex usage interface, while Codex users can check status information from within the CLI. However, several developers are asking for more detailed task-level explanations showing exactly what caused their allowance to decrease.
That demand has become more noticeable as reports of discrepancies have appeared. In one GitHub issue, a Plus user said the account showed 58% of the weekly quota remaining while Codex simultaneously reported that the usage limit had been reached.
Other reports describe similar inconsistencies between displayed usage and backend information. One Pro user reported that the usage panel showed only 1% remaining while the underlying usage endpoint reportedly indicated 19% remained.
Reports Extend Beyond a Single Account
The complaints are not limited to one subscription level or one type of workload.
Another developer reported that weekly usage moved from 0% used to 97% used within a single day, making it difficult to determine whether the problem involved actual consumption or usage-metering behavior.
A separate August report claimed that the weekly usage indicator dropped sharply even though the user had done very little intentional work. Another user said their remaining allowance continued declining even while Codex was closed.
There are also reports of reset-related problems. In one case, a developer said their usage fell from 86% remaining to 1% and the scheduled reset time unexpectedly changed after a manual reset.
These reports point to several possible explanations, including complex workloads, token accounting, delayed reconciliation, subagent activity or problems with how usage information is displayed. The available reports alone do not establish which explanation is responsible for every case.
Why Codex Usage Can Vary
Codex is not charged simply according to the number of prompts a user sends. OpenAI’s current rate card uses token-based accounting, which means the amount of input and output processed can affect consumption. Cached input is also included in the calculation.
A coding task involving a large project context can therefore consume considerably more resources than a short request. Long-running sessions, complex reasoning and multiple agents can also make usage harder for users to estimate.
This is particularly important for professional developers who use Codex continuously. If the usage meter changes substantially during a long engineering session, users need enough information to understand whether the consumption is expected or the result of an accounting or display problem.
OpenAI Has Not Confirmed a Secret Quota Reduction
At present, there is no clear evidence in the reports reviewed that OpenAI has intentionally cut Codex allowances by 50%.
What is clear is that multiple users have reported unusually rapid consumption and inconsistencies involving Codex usage indicators. Several of these reports remain open on the official Codex repository, meaning the issues should not automatically be treated as resolved or confirmed policy changes.
The distinction matters because a genuine reduction in plan allowances would be different from a usage-metering bug, a change in token accounting, delayed usage reporting or a mismatch between the user interface and backend data.
For now, developers concerned about their Codex consumption can monitor the official usage information and compare it with their local activity and token records where available.
The growing number of reports nevertheless highlights a broader issue for AI coding platforms: users need predictable limits and transparent accounting. As developers increasingly rely on AI for professional software engineering, understanding exactly how usage is calculated can be almost as important as the model’s coding performance.
