GPT-5.3-Codex significantly upgrades your coding and professional workflows.
Available today in all Codex Surfaces: Codex app, CLI, IDE extensions, and web
25% Faster Performance: Accelerate your coding and agentic tasks with improved inference speed.
Interactive Agentic Coding: Steer your agent in real-time during complex, multi-file tasks without losing context.
Superior Multi-language Coding: Top performance across multiple languages, surpassing GPT-5.2-Codex on SWE-Bench Pro.
Robust Terminal & OS Interaction: Improved accuracy on Terminal-Bench 2.0 and OSWorld-Verified for seamless system-level tasks.
Better Web & Frontend Development: Automatic, production-quality website creation from simple prompts with improved UX/UI.
Advanced Cybersecurity: Direct vulnerability detection and strengthened safeguards, earning our highest cybersecurity capability rating.
Broader Professional Workflows: Enhanced productivity in tasks like documentation, presentations, data analysis, and structured outputs
I’m loving the coding, but not the speed. I don’t know if it’s a load issue, but it’s probably 3x slower than 5.2 codex in my experience so far today. Painfully slow. I’m hoping this is just a launch day issue.
As @vb stated, API access has not yet been released so I assume prices are TBD too.. Auth via ChatGPT for the time being and the price you pay is your monthly sub.
I’ve been building this, which I think is quite appropriate, all things considered.
(FYI works on many Discourse instances, but not yet on this site as this site blocks automated remote username password interaction - the malicious bots always spoil the fun )
No luck getting the VSCode latest “release” or “preview” extension logged in via the “Sign in with ChatGPT” flow. It launches a browser, wanting your login, password, (then one-factor on a mobile device that doesn’t even support ChatGPT app any more, then needing “try with email”) by email login code, then simply fails. Tried having it launch OS default browser Firefox, Chromium, and Edge - all fail.
Gone as far as reinstalling Codex extension, re-updating VSCode, reinstalling WSL with defaults, renaming the .codex directory, windows sandbox or WSL mode, restarting the whole machine..
“Device code” offers no opportunity nor difference in the auth failure. Whatever “remote browser data” is has the setting enabled (scarfing up browser info into training?).
There is absolutely no troubleshooting info for IDE extensions in any docs beyond, “Sign in with your ChatGPT account or an API key to get started.”
edit - finally got this to go through a day and a half later for no particular change made.
Logging in with API key still has a tooltip “Powered by your ChatGPT account.”, BTW
I’m a ChatGPT Plus subscriber, but I can’t seem to find the GPT-5.3-Codex model in my model picker/dropdown menu.
Is this model currently being rolled out gradually, or should it be available to all Plus users immediately? I’ve tried refreshing and logging out/in, but it still doesn’t appear.
Based on my testing so far, yes, it is faster, but I would argue that it is neither better nor smarter than version 5.2 yet. I’m getting mixed results. In some areas, the approach is better, but not necessarily much better, while in others, it goes down some fairly strange routes. One of these general issues is trying to solve bugs with fallbacks instead of fixing the issue itself (i.e. migrating around it). But let’s wait and see; it’s still early days. In any case, the advancements in the last 6 months have been amazing. Just as info.
I think a lot of people forget how fast the industry is moving when you really think about it. ChatGPT had a lot of users super quick. The scaling of things now is crazy. I think we’ll begin to see the models improving themselves soon … then maybe another hockey-stick moment? I believe so, but we shall see.
Yeah, though, we’ve become so accustomed to new, better, faster, cheaper, I think we sometimes forget how fast everything is moving. There’s not enough electricity to really keep brute-forcing it yet, so I’m wondering if there will be other breakthroughs that will accelerate it even more.
Interesting thread. Thanks to everyone participating!
I cannot answer - I have a free account, installed a brand new VSCode on a new computer image with the offered extension for Codex - by OpenAI, logged in via ChatGPT. Contrary to first message cut-and-paste “available today in all Codex surfaces”, but in agreement with its “available today for all paid ChatGPT users”, it is not available in this surface pre-populated into the models listing.