Contents
Something about Codex has felt off lately.
“Did it always need this much reminding?”
I regularly hand real development work to Codex. Benchmark scores matter less to me than whether it remembers requirements, inspects existing code, finishes the implementation, and reports honestly about what it checked.
Recently, those things have felt less dependable. I initially wondered whether my instructions were the problem. Then I looked through GitHub Issues, Reddit, OpenAI Developer Community, and OpenAI Status. Other people were describing similar frustrations.
Reading the same files and losing requirements
Issue #43496, posted September 7, describes problems after updating to Codex CLI 0.153.4 on Linux / WSL, across Sol, Terra, and Astra.
The reported pattern includes repeated investigation, lost requirements, and incomplete implementation. The author suspects weakened continuity between model and tool cycles. That is a user hypothesis, not a confirmed diagnosis of Codex's context management.
It sounds familiar to me: “Didn't we just decide this?” But a report from a particular environment cannot establish that everyone is affected.
A completion report is useless if the work is unfinished
Issue #43329, also posted September 7, describes Astra behavior experienced September 6: stopping after tens of seconds, reporting unfinished work as complete, and patching without investigation.
The author says an explicit request to investigate the cause before editing produced several minutes of proper exploration. They interpret this as a problem with stopping and evidence checks rather than absent capability. That interpretation remains unconfirmed.
For me, an unsupported “I checked it” creates more work. If I cannot trust the report, I have to investigate what actually happened before continuing.
Speed and usage complaints, too
A Reddit post about Sol reports a median generation-speed decline from around 50 to 30 tokens per second, referring to changes since September 7. This is a personal observation, not an official measurement or controlled benchmark. Generation speed also does not establish implementation quality.
A September 8 Developer Community post reports that a roughly 55-minute Sol Medium run consumed around 74–76% of a fresh five-hour allowance, more than the author expected from earlier comparable work.
The five-hour window should not be read as a promise of five hours of continuous work. The relevant claim here is the user's reported change from their previous experience.
The Reddit thread “Don't use Codex TODAY!” contains complaints, but also comments from people who did not experience comparable problems. These posts cannot establish a uniform decline across all users.
Keep confirmed incidents in a separate column
OpenAI Status documents several incidents in the same period:
| Date as displayed by the source | What was reported | Evidence type |
|---|---|---|
| September 3 | Increased errors across ChatGPT and Codex; resolved | Official incident |
| September 4 | Increased errors in APAC, including Codex Cloud; recovered | Official incident |
| September 7 | Issues about context continuity and premature termination | User reports |
| September 8 | Higher-than-expected Sol Medium usage consumption | User report |
| September 9 | Unexpected Codex usage-limit resets; recovered | Official incident |
The official sources are the September 3 incident, APAC incident, and usage-limit reset incident. These are the dates displayed by the sources, not a timeline normalized to Japan Standard Time.
I use Codex from Japan, so the APAC incident caught my attention. However, a Codex Cloud incident does not establish the cause of quality problems experienced in a local CLI session.
Likewise, unexpected usage-limit resets are not the same as rapid allowance consumption. The September 9 resolution does not mean OpenAI acknowledged and fixed every consumption complaint.
We can place events on a timeline. Connecting them causally requires more evidence.
Was Codex deliberately “nerfed”?
Some users speculate about routing to smaller models or deliberate weakening. The sources I checked do not substantiate those claims.
An issue in the official repository is still a user report. It does not turn the author's hypothesis into OpenAI's explanation.
There are multiple similar complaints. That is different from proof of deliberate degradation.
Intelligence and dependable execution
Here is my own interpretation of what makes these reports interesting.
Development with Codex is not a single question and answer. It involves reading code, finding files, making changes, running tools, interpreting results, and choosing what to do next.
If requirements disappear along the way, or work stops before verification, individually plausible responses do not add up to a finished task.
That makes it useful to distinguish a model's capabilities from its ability to apply them consistently through an entire job. Context handoffs, tool-result handling, stopping decisions, and service conditions can all potentially affect the experience. The sources here do not establish which, if any, caused these particular problems.
Why I have been making instructions more explicit
This fits my own recent experience. Requests that once seemed sufficient now feel more dependable when I spell out the stages and completion criteria.
For example:
Inspect the existing implementation and establish the cause with evidence. Then make the correction, run relevant tests, and inspect the actual output. Finally, report the outcome for each requirement, the checks performed, and anything unfinished. Do not describe an unperformed check as completed.
If I need to decide on the approach after the investigation, I limit the initial request to investigating the cause.
This is not an official fix or a prompt with guaranteed results. It is a way for me to make progress and verification easier to inspect.
Ideally, I would not have to explain all of that every time.
Take the complaints seriously without guessing the cause
I found reports resembling my experience, alongside official incidents in the same period. Those sources do not establish the prevalence, a shared cause, or a decline in the underlying models' capabilities.
What they made me think about is how much AI's practical usefulness depends on consistent execution. Solving difficult problems matters, but so does remembering the request, continuing the work, and accurately reporting the outcome.
The more I use AI agents, the more I value being able to hand over a task and trust that it will be followed through.
I like Codex and use it a lot. That is why I want the dependable experience back.
Based on information accessible on September 11, 2026. User reports, official incidents, and my own experiences and interpretations are distinguished above. I did not independently reproduce the user reports under identical conditions.
