There is a strange failure mode that I have been noticing in OpenAI models: in a longer session, the model can suddenly answer an earlier prompt instead of the one right in front of it.
You can be done with topic A, move on to something else (B), and then suddenly get a reply that clearly belongs to A. Sometimes the model even handles the new topic correctly at first, which makes the jump back even more confusing.
This seems to be not just a weak answer or a misunderstanding. It looks more like the model briefly loses track of which part of the conversation it is supposed to be responding to.
The problem is not only that the answer is wrong. The problem is that it appears to be aimed at the wrong place in the conversation.
What other people report
Based on issue reports and discussions from others around this, the problem does seem to be real. Multiple people have described the same general behavior: older prompts resurfacing, stale answers appearing in the middle of a newer discussion, and the model drifting back to a previous part of the conversation.
I found most reports are Codex-related, but that does not necessarily mean the problem belongs only there. Those environments simply make the issue easier to notice because people are working through longer, more structured sessions and can tell more clearly when a reply belongs to an earlier step.
So I don't think it is solely a Codex problem and also not a coding-only problem. It is a conversation-tracking problem that becomes especially visible in longer and more demanding sessions.
Where the problem likely sits
There is still no clean public answer on the root cause.
The reports do not point to one single broken CLI version. They also do not support the idea that one model family is completely unaffected. Similar complaints show up across several GPT-5.x variants, including both -codex and non--codex models.
So this does not look like one bad release, one bad setting, or one bad wrapper. It looks more like an underlying model behavior that becomes easier to trigger in certain environments.
What to do about it
There is no clear public fix yet.
You can obviously keep sessions shorter, so split unrelated topics into separate threads.
The larger point
The interesting part of this issue is not that the model gets things wrong. Everyone already knows that models get things wrong.
What stands out here is the type of mistake. When a model answers an older prompt while presenting the reply as if it were about the current one, this feels just totally off and makes you loose all trust in the model as continuity is one of the basic things people expect.
Links
Here are the main links that support the points above:
- Codex replies to earlier messages instead of latest one — GitHub issue #8648
- Context corruption / attention issue — answers older question — GitHub issue #14408
- Stale prior response after tool failure — GitHub issue #14253
- Cross-session or cross-worktree leakage — GitHub issue #15027
- GPT-5.4-era unrelated context / stale behavior report — GitHub issue #13823
- Repeated-problem / loss-of-place behavior with GPT-5.2 — GitHub issue #11897
- gpt-5.3-codex context failure report — GitHub issue #14772
- Model-routing confusion involving gpt-5.3-codex — GitHub issue #11867
- Resume/new-session workaround discussion — GitHub issue #8256
- Comment suggesting this looks more like model behavior than client harness — GitHub issue #11527
- Codex changelog: GPT-5.4 rollout in Codex CLI
- Codex changelog
- Codex model config showing available model options
Subscribe to my Newsletter
Get the latest updates delivered straight to your inbox
I respect your privacy. Unsubscribe at any time.