What version of Codex CLI is running?
0.145.0
What subscription do you have?
ChatGPT Pro ($200/month)
Which model were you using?
gpt-5.6-sol ultra
What platform is your computer?
Linux 6.6.87.2-microsoft-standard-WSL2 x86_64 (WSL2, Debian, Windows host)
What terminal emulator and version are you using (if applicable)?
Windows Terminal 1.24.11911.0 running WSL2/Debian
Codex doctor report
overallStatus: ok
codexVersion: 0.145.0 (latest available: 0.146.0)
model: gpt-5.6-sol
model provider: openai
auth mode: chatgpt
platform: linux-x86_64 (WSL2, Debian 13.0.0 trixie)
network/provider reachability: ok
All checks passed — no environment, auth, network, or config issues detected.
Full JSON output available on request.
What issue are you seeing?
I'm writing this as a Pro user paying $200/month and as someone with over 15 years of professional full-stack development experience, also working with deep learning models on a regular basis, because I think this deserves more than a vague "it feels worse" complaint. I want to document this properly.
Since yesterday, the quality of code generation has collapsed to a level that is genuinely hard to believe coming from what's supposed to be the top tier model. I'm not talking about small inconsistencies or a model having an off moment on a tricky prompt. I'm talking about basic, previously reliable tasks failing outright, repeatedly, across completely unrelated parts of my stack.
Beyond specific bugs (detailed in "steps to reproduce" below), the general frontend judgment has taken a hit too. Layout decisions, component structure, the kind of architectural reasoning that used to be handled competently is now inconsistent and careless in a way that feels categorically different, not just slightly worse.
I want to explain why I don't think this is a normal fluctuation, because I know that's the first assumption people make. I work with LLMs professionally, so I understand what normal variance looks like: occasional stylistic differences, edge cases handled inconsistently, prompts needing slight rewording. What I'm describing here is not that. This is a simultaneous drop across multiple, completely unrelated task types, appearing from one day to the next, with no change on my end to prompting style, project structure, or anything else. Based on what I'm observing, my working hypothesis is that this pattern points toward something changing on your side, whether that's a quiet model swap, a quantized or distilled variant being routed to certain users or under certain load conditions, or some other backend decision that wasn't communicated. To be clear, this is a hypothesis on my part, not a claim of fact, and it's exactly why I'm asking for a direct answer below instead of assuming.
What steps can reproduce the bug?
Two concrete, reproducible cases from today, both on model gpt-5.6-sol (reasoning effort: Ultra):
-
Asked for a simple section component in Next.js using React, something this tool used to handle without any issue, and the output came back broken and incomplete. Not stylistically different, actually non-functional.
-
Styled a single button in SCSS and the change introduced regressions in selectors elsewhere in the stylesheet that had absolutely nothing to do with the request. That kind of scope leakage, where the model touches things outside the boundaries of what was asked, is the type of mistake I'd expect from a much smaller or less capable model, not from something marketed as your flagship offering.
Environment: Codex CLI 0.145.0, WSL2/Debian 13.0.0, Windows Terminal, model gpt-5.6-sol. Full codex doctor output attached below confirms no environment, auth, network, or config issues — this is not a local setup problem.
What is the expected behavior?
Code generation quality consistent with the advertised top-tier model ("ChatGPT 5.6 Ultra" / gpt-5.6-sol, reasoning effort Ultra), matching the reliability I experienced with this same setup until yesterday: functional, scoped output for basic Next.js/React and SCSS tasks, without unrelated regressions or broken components.
Additional information
So here's what I actually want answered, and I'd like a specific response, not a generic support macro:
Did anything change in terms of which model or variant is serving requests over the last 24 to 48 hours? A simple yes or no would already tell me a lot.
If there are capacity constraints causing Pro tier traffic to be served by a reduced capability model at certain times or under certain conditions, that needs to be disclosed openly. I'm paying for a specific tier and I expect to know what I'm actually getting, not to find out through degraded output that something quietly changed underneath me.
If this turns out to be a confirmed bug or regression that has nothing to do with any intentional change, I want an actual timeline, not "we're aware and looking into it," which by now feels like a placeholder answer rather than a real update.
To be clear about the stakes here, I use this tool for production level client work. At the quality level I'm seeing right now, it simply isn't usable for that, and that makes it very difficult to justify what I'm paying every month. I'd genuinely prefer to keep working with this tool, I've built a good part of my workflow around it, but staying silent on something this measurable and this sudden is exactly what pushes experienced users toward alternatives that are currently performing better for a fraction of the cost.
I'm not looking for sympathy here, I'm looking for a direct, specific, and honest answer about what actually changed.
What version of Codex CLI is running?
0.145.0
What subscription do you have?
ChatGPT Pro ($200/month)
Which model were you using?
gpt-5.6-sol ultra
What platform is your computer?
Linux 6.6.87.2-microsoft-standard-WSL2 x86_64 (WSL2, Debian, Windows host)
What terminal emulator and version are you using (if applicable)?
Windows Terminal 1.24.11911.0 running WSL2/Debian
Codex doctor report
What issue are you seeing?
I'm writing this as a Pro user paying $200/month and as someone with over 15 years of professional full-stack development experience, also working with deep learning models on a regular basis, because I think this deserves more than a vague "it feels worse" complaint. I want to document this properly.
Since yesterday, the quality of code generation has collapsed to a level that is genuinely hard to believe coming from what's supposed to be the top tier model. I'm not talking about small inconsistencies or a model having an off moment on a tricky prompt. I'm talking about basic, previously reliable tasks failing outright, repeatedly, across completely unrelated parts of my stack.
Beyond specific bugs (detailed in "steps to reproduce" below), the general frontend judgment has taken a hit too. Layout decisions, component structure, the kind of architectural reasoning that used to be handled competently is now inconsistent and careless in a way that feels categorically different, not just slightly worse.
I want to explain why I don't think this is a normal fluctuation, because I know that's the first assumption people make. I work with LLMs professionally, so I understand what normal variance looks like: occasional stylistic differences, edge cases handled inconsistently, prompts needing slight rewording. What I'm describing here is not that. This is a simultaneous drop across multiple, completely unrelated task types, appearing from one day to the next, with no change on my end to prompting style, project structure, or anything else. Based on what I'm observing, my working hypothesis is that this pattern points toward something changing on your side, whether that's a quiet model swap, a quantized or distilled variant being routed to certain users or under certain load conditions, or some other backend decision that wasn't communicated. To be clear, this is a hypothesis on my part, not a claim of fact, and it's exactly why I'm asking for a direct answer below instead of assuming.
What steps can reproduce the bug?
Two concrete, reproducible cases from today, both on model gpt-5.6-sol (reasoning effort: Ultra):
Asked for a simple section component in Next.js using React, something this tool used to handle without any issue, and the output came back broken and incomplete. Not stylistically different, actually non-functional.
Styled a single button in SCSS and the change introduced regressions in selectors elsewhere in the stylesheet that had absolutely nothing to do with the request. That kind of scope leakage, where the model touches things outside the boundaries of what was asked, is the type of mistake I'd expect from a much smaller or less capable model, not from something marketed as your flagship offering.
Environment: Codex CLI 0.145.0, WSL2/Debian 13.0.0, Windows Terminal, model gpt-5.6-sol. Full codex doctor output attached below confirms no environment, auth, network, or config issues — this is not a local setup problem.
What is the expected behavior?
Code generation quality consistent with the advertised top-tier model ("ChatGPT 5.6 Ultra" / gpt-5.6-sol, reasoning effort Ultra), matching the reliability I experienced with this same setup until yesterday: functional, scoped output for basic Next.js/React and SCSS tasks, without unrelated regressions or broken components.
Additional information
So here's what I actually want answered, and I'd like a specific response, not a generic support macro:
Did anything change in terms of which model or variant is serving requests over the last 24 to 48 hours? A simple yes or no would already tell me a lot.
If there are capacity constraints causing Pro tier traffic to be served by a reduced capability model at certain times or under certain conditions, that needs to be disclosed openly. I'm paying for a specific tier and I expect to know what I'm actually getting, not to find out through degraded output that something quietly changed underneath me.
If this turns out to be a confirmed bug or regression that has nothing to do with any intentional change, I want an actual timeline, not "we're aware and looking into it," which by now feels like a placeholder answer rather than a real update.
To be clear about the stakes here, I use this tool for production level client work. At the quality level I'm seeing right now, it simply isn't usable for that, and that makes it very difficult to justify what I'm paying every month. I'd genuinely prefer to keep working with this tool, I've built a good part of my workflow around it, but staying silent on something this measurable and this sudden is exactly what pushes experienced users toward alternatives that are currently performing better for a fraction of the cost.
I'm not looking for sympathy here, I'm looking for a direct, specific, and honest answer about what actually changed.