Severe Performance Degradation in GPT-5.3-Codex - Model Effectively Crippled

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #12225

Powered by Codex Action

+1 I just popped in to codex repo to see if anyone else has reported Codex issues. For me it has degraded to the point where it's basically unusable.

I noticed that Tibo recently tweeted on X about "amazing things coming for GPT-5.3-Codex over the next 10 weeks." I'm genuinely surprised—this isn't exactly the kind of "amazing" I expected. Could someone clarify the roadmap?

Is it going to be something like this?

• Week 1: Significant reduction in performance (as if the model underwent a "frontal lobotomy").
• Week 2: Restoration to original performance (so users who've adapted to the severely reduced performance feel an illusionary "improvement").

Could it be that the first two weeks were intentionally designed to set the stage for easy wins later?

Clarification on what's really happening would be greatly appreciated.

I'm not aware of any changes we've made to the gpt-5.3-codex model or its hosting environment in the past week. Are you sure that you haven't made any configuration changes in the client (e.g. the reasoning effort setting)? Are you using any non-default config settings in config.toml?

If you observe this behavior again, please use the /feedback slash command to upload your logs and session details so we can investigate further.

@etraut-openai, Upon further investigation, I've observed that the performance degradation specifically occurs when using exec commands or running via the app server v2.
Interactions via the CLI environment seem unaffected.

This is particularly concerning when contrasted with Sam Altman's recent acquisition of OpenClaw.
On the surface, OpenAI claims to be making Codex more "Open" and externally accessible,
yet internally, the deliberate performance throttling suggests a "Closed" approach behind the scenes.

@ehgus8, you said that you're seeing a performance degradation when using exec or app server. Can you be more specific? Do you have any non-default configuration settings defined in your config.toml file? If you see a behavior that you think looks off, please use the /feedback slash command to upload your logs and session details and post the thread ID so we can investigate further.

@etraut-openai, Thanks for the follow-up. Here are my current config.toml
settings that may be non-default:

• model = "gpt-5.3-codex"
• model_reasoning_effort = "xhigh"
• personality = "pragmatic"
• [notice] hide_full_access_warning = true
• [tui] status_line = ["model-with-reasoning", "current-dir", "git-branch",

"five-hour-limit", "context-remaining"]

• trust_level = "trusted" is set for my local project paths

Feeling the same in the last two weeks, some degradation is happening, and I have the Pro plan.

I stopped using Plan Mode because we agreed the plan and when switch to execute mode to implement, Codex ended up implementing totally in a different way (even it works in the end, but with bad code, no following the agreements we did in the plan)

I've asked twice above for session details to be uploaded via the /feedback slash command. Without this, there's little or nothing I can do to investigate this further.

We haven't made any changes to the gpt-5.3-codex model since it was released. If you think you're experiencing a difference in behavior, that's likely caused by some configuration change or a change in how you're using the model (e.g. different problem type, different prompting style, different thread/context length, different configuration settings, etc.).