Significant model quality regression in codex-5-5 (xhigh) — Instructions not followed, repeated regressions, affects both CLI and IDE
Open 💬 10 comments Opened May 26, 2026 by AndyAlvarezDev
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What issue are you seeing?
Since a recent update, the model behavior behind the xhigh effort level (which we assume maps to codex-5-5 or equivalent) has degraded drastically compared to what it was a few weeks ago. The quality drop is so significant that it no longer feels like the same model — it resembles a heavily quantized or distilled variant rather than the full-capability model we were previously using.
Observed symptoms:
- The model frequently ignores explicit instructions in
AGENTS.mdand in prompt context, even when they are clear, concise, and well-structured. - It produces repeated regressions: changes that were already reviewed and corrected are reintroduced in subsequent turns.
- It fails to maintain task coherence across multi-step workflows — it loses track of context and reverts to generic or incorrect behavior mid-task.
- It appears to hallucinate tool usage or skip tool calls it should be making.
- Overall reasoning quality feels noticeably inferior to previous versions.
This is not an isolated or edge-case scenario. It makes the tool practically unusable for serious development workflows.
Questions for the team
- Was the underlying model recently updated, swapped, or re-quantized? The behavioral delta is large enough that it does not feel like the same checkpoint.
- When can we expect a quality improvement or rollback? Given that this affects both CLI and IDE users, it would be helpful to have a public timeline or at least an acknowledgment that the regression is known.
Environment
- Codex CLI: latest
- IDE: Codex (latest)
- OS: Linux / Windows
- Effort level:
xhigh
What steps can reproduce the bug?
- Set effort level to
xhigh. - Provide a detailed
AGENTS.mdwith explicit coding conventions and task instructions. - Ask the model to implement a non-trivial feature (e.g., add a validated form field with specific error handling logic).
- Observe that the model either ignores the conventions, introduces regressions from prior turns, or produces code that contradicts the stated requirements.
- Repeat across multiple sessions — the behavior is consistent, not random.
Reproducible in:
- Codex CLI (latest version)
- IDE integration (VS Code extension)
This is not a CLI-only issue. The same degraded model behavior is present in the IDE, which rules out any client-side regression.
What is the expected behavior?
_No response_
Additional information
_No response_
10 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
I got the same problem in these few days!!
I'm saying this with full authority as a senior developer with 15+ years of experience, with a strong background in machine learning — this is not a beginner complaint.
I'm on a Pro plan and I spend my entire day iterating just to get something working, because every fix introduces a new regression. That's not a workflow, that's a treadmill. The model has become genuinely unusable for serious work.
And I'm not alone. Go check Reddit — the community is full of people reporting the exact same thing right now. When experienced engineers are struggling to get consistent results from a coding agent, imagine what it's like for someone with less technical background.
I hate to say it, but at this point some open-source Chinese models are performing more reliably on instruction-following tasks than whatever is running behind xhigh right now. That should be a wake-up call.
Please treat this as a priority. The product reputation is at stake.
I encountered simular problems!
Got the same trouble
It's been several days and this issue still has no assignee, no acknowledgment, nothing. Just silence.
I'm paying $200/month for Pro and the tool is basically broken for any serious workflow. The model at xhigh ignores instructions, keeps reintroducing changes that were already corrected, and loses track of context mid-task. This isn't a minor glitch — it's a fundamental regression that makes Codex unreliable for real work.
The least we deserve at this price point is someone from the team saying "we know, we're on it." That's it. Not a fix, not a timeline — just basic acknowledgment that you're aware this is happening.
Is anyone actually looking at this?
Is this suspected of commercial fraud?
Nobody is looking. Nobody gives a shit. And the issue has gotten worse.
I want to add a stronger signal here as a heavy Codex user.
My blocker is not token usage, latency, or cost. My blocker is intelligence. I would gladly spend substantially more tokens for a true maximum-intelligence mode if it can produce better engineering work.
Right now, GPT-5.5 xhigh in Codex feels materially weaker than Claude's strongest coding/reasoning mode for complex debugging, long-context planning, code review, and multi-step implementation. The failure mode is not just "slower output"; it is shallow fixes, losing task coherence, missing constraints already present in AGENTS.md or the prompt, and stopping at plausible answers without enough verification.
Please prioritize a real maximum-intelligence / deep-work mode for Codex, even if it is expensive, slower, or explicitly opt-in. Some users do not want cheaper tokens. We want stronger reasoning.
There's a lot of it about (in my experience, now unusable very suddenly, post July 13th updates #33070)
Here's my 'favourite' reply from the Codex App so far to date, today: