It sounds like it can make simple tasks much more correct. It's impressive to me. Today coding agent tends to pretend they're working hard by generating lots of unnecessary code. Hope it's true
I'm happy with medium reasoning. My projects have been in Go, Typescript, React Dockerfiles stuff like that. The code almost always works, it's usually not "Clean code" though.
SWE-bench performance is similar to normal gpt-5, so it seems the main delta with `gpt-5-codex` is on code refactors (via internal refactor benchmark 33.9% -> 51.3%).
As someone who recently used Codex CLI (`gpt-5-high`) to do a relatively large refactor (multiple internal libs to dedicated packages), I kept running into bugs introduced when the model would delete a file and then rewrite in (missing crucial or important details). My approach would have been to just the copy the file over and then make package-specific changes, so maybe better tool calling is at play here.
Additionally, they claim the new model is more steerable (both with AGENTS.md and generally). In my experience, Codex CLI w/gpt-5 is already a lot more steerable than Claude Code, but any improvements are welcome!
[0]https://github.com/openai/codex/blob/main/codex-rs/core/gpt_...
[1]https://github.com/openai/codex/blob/main/codex-rs/core/prom...
GPT-5 may underwhelm with the same sparse prompt, as it seems to do exactly what's asked, not more
You can still "fully vibe" with GPT-5, but the pattern works better in two steps:
1. Plan (iterate on high-level spec/PRD, split into actions)
2. Build (work through plans)
Splitting the context here is important, as any LLM will perform worse as the context gets more polluted.
SWE-bench is a great eval, but it's very narrow. Two models can have the same SWE-bench scores but very different user experiences.
Here's a nice thread on X about the things that SWE-bench doesn't measure:
Prompting is different, but in a good way.
With Claude Code, you can use less prompting, and Claude will get token happy and expand on your request. Great for greenfield/vibing, bad for iterating on existing projects.
With Codex CLI, GPT-5 seems to handle instructions much more precisely. It won't just go off on it's own and do a bunch of work, it will do what you ask.
I've found that being more specific up-front gets better results with GPT-5, whereas with Claude, being more specific doesn't necessarily stop the eagerness of it's output.
As with all LLMs, you can't compare apples to oranges, so to clarify, my experiences are primarily with Typescript and Rust codebases.
It seems about half my sessions quickly become "why did you do that? rip __ out and just do ___". Then again, most of the other sessions involve Codex correctly inferring what I wanted without having to be so specific.
Beyond that, purely anecdotal and subjective, but this model does seem to do extensive refactors with semi precise step-by-step guidance a bit faster (comparing GPT-5 Thinking (Medium) and GPT-5 Codex (Medium)), though adherence to prompts seems roughly equivalent between the two as of now. In any case, I really feel they should consider a more nuanced naming convention.
New Claude Sonnet 3.7 was a bit of a blunder, but overall, Anthropic has their marketing in tight order compared to OpenAI. Claude Code, Sonnet, Opus, those are great, clear differentiating names.
Codex meanwhile can mean anything from a service for code reviews with Github integration to a series of dedicated models going back to 2021.
Also, while I do enjoy the ChatGPT app integration for quick on-the-go work made easier with a Clicks keyboard, I am getting more annoyed by the drift between Codex VSCode, Codex Website and Codex in the ChatGPT mobile app. The Website has a very helpful Ask button, which can also be used to launch subtasks via prompts written by the model, but such a button is not present in the VSCode plugin, despite subtasks being something you can launch from the VSCode plugin if you have used Ask via the website first. Meanwhile, the iOS app has no Ask button and no sub task support and neither the app, nor VSCode plugin show remote work done beyond abbreviations, whereas the web page does show everything. Then there are the differences between local and remote via VSCode and the CLI, ... To people not using Codex, this must sound insane and barely understandable, but it seems that is the outcome of spreading yourself across so many fields. CLI, dedicated models, VSCode plugin, mobile app, code review, web page, some like Anthropic only work on one or two, others like Augment three, but no one else does that much, for better and worse.
I like using Codex, but it is a mess with such massive potential that needs a dedicated team lead whose only focus is to untangle this mess, before adding more features. Alternatively, maybe interview a few power user on their actual day to day experience, those that aren't just in one, but are using multiple or all parts of Codex. There is a lot of insight to be gained from someone who has an overview off the entire product stack, I think. Sending out a questionnaire to top users would be a good start, I'd definitely answer.
It is inevitable.
And before someone says it, I do happen to have my own codex like environment complete with development containers, browser, github integration, etc.
And I'm happy to pay a mint for access to the best models.
>For developers using Codex CLI via API key, we plan to make GPT‑5-Codex available in the API soon.
Ditched my Claude code max sub for the ChatGPT pro $200 plan. So much faster, and not hit any limits yet.
The $200 pro feels good value personally.
No really Jokes apart, ChatGPT and Trump are awesome, all works and keeps getting better by the day.
Both were struggling yesterday, with Claude being a bit ahead. Their biggest problems came with being "creative" (their solutions were pretty "stock"), and they had trouble making the simulation.
Tried the same problem on Codex today. The design it came up with still felt a bit lackluster, but it did _a lot_ better on the simulation.
I’m imagining if it can navigate the codebase and modify tests - like add new cases or break the tests by changing a few lines. This can actually verify if the tests were doing actual assertions and being useful.
Thorough reviewing like this probably benefits me the most - more than AI assisted development.
For people that have not tried it in say ~1 month, give Codex CLI a try.
incomingpain•4h ago
NitpickLawyer•1h ago