I've been using 5.3 Codex for 3 weeks. It's an incredible model. I've built so much stuff with it. Made a vid showing everything I love about it, as well as a few call-outs of things I hope OpenAI changes.

been splitting between opus and codex based on the task. codex is cleaner for isolated function-level work, opus handles multi-file refactors better. which use cases blew you away most?

how did openai pick you for early access? Because i have never seen a builder who has early access to openai models said sth negative about it. For example Matt Schumer shared an article about codex and it was gpt 5.3 vs opus 4.5. I think this article had already been written before that launch . I guess opus 4.6 was unexpected for gpt 5.3 early users and for open ai. They launched it right after opus 4.6. This has been same since GPT 5. Gpt 5 was not even close to Claude though. it is just weird.

I talked a decent bit of shit on 5.1, didn't feel like a meaningful upgrade. I even called out a bunch of things I don't like at the end of this video. @benhylak also often has early access and was VERY unhappy with gpt-5.2. Anthropic doesn't do much early access afaik

@nurullah_kuus getting well-known people to mindlessly shill your stuff is not a repeated game. openai gets this, and want people who are honest so they can get a shot at addressing their feedback before they release something.

Yeah, but it seems like an ad from an outside viewer. Especially when they say stuff like “I tried it for 3 weeks and this is goat mode.” It clearly wasn’t until 5.3. They should do that too btw to get feedback, but being a bit more objective would be nice. Your reviews seem more objective though.

It’s a really cool model! I built using codex 5.3 with less than 25% of laptop charge left 😭😭

thank you, 5.3

the planning gap is the sleeper issue. codex nails isolated tasks but the moment you need it to hold a mental model of your whole codebase across multiple files, opus pulls ahead. i run both daily and the pattern that stuck: codex for speed runs on well scoped tickets, opus when the agent needs to reason about dependencies nobody explicitly told it about. context management is doing more heavy lifting than raw model capability at this point.

They think they're evolving ("hybrid thing through recursion"), but they're just looping in the same old deception: creation worshiping itself (Romans 1:25). No trembling before the Creator. Just building mirrors and calling it symbiosis.

Once you plan workflows and apply tighter prompt structure, Codex really shines.

This is one of my favs of your Theo, thanks for the deep dive.

The improvements in reasoning make complex tasks feel more manageable, though context limits can still trip things up occasionally.

does codex cli works the same as the desktop app with 5.3 (so you can use it on linux?)

what do you think?

The shill is obvious now

When builders start asking for changes instead of alternatives, you’re onto something.

5.3 Codex sounds like a game changer Hearing what you love and what could improve is super useful for anyone thinking of building with it Excited to see your video breakdown

Well I guess getting off Sam can be considered a job too if you’re unemployed… #keep4o

no lies found

It has been thoroughly enjoying following your journey of AI-augmented development. As models become smarter, tech-savvy individuals would fear losing their leverage. However, you didn’t, and you’re doing great!

Yeah, it’s crazy, been using it a lot, I need it to get better at planning out code/architecture so I can get rid of Claude. Maybe it’s just me and I’m not optimizing prompts correctly but that’s one of the few places I have found it to lack a little

They have already lobotomised the model. At least in terms of speed, it is way slower than at the launch.

Banger ! Really great and interesting video

As a Cursor Ultra subscriber, I was reluctant to try Codex. But after a week with 5.3 and a few days with Codex proper... I'm convinced the future of dev is less about the specific code and more about agent orchestration. Painful to admit. But here we are.

@theo Since you're a heavy #codex user like myself, I thought you might like this. It's a theme customizer for the Codex desktop application 📥

🥰

Interesting

@theo, the Codex model indeed allows one to build complex applications fast, and here's a helpful guide.

Sharing both strengths and pain points really helps others understand its real-world potential.

Did you manage to write tests with it eventually?

Context length could use more expansion.

theo this was great! learned a lot. thanks for sharing. ✌🏼

I need shoe 👟 today theo

@SuperGrok @grok @AskPerplexity Summarize this video for me.

Codex 5.3 is a smart colleague

What’s fascinating about Codex 5.3 isn’t just raw code generation — it’s iterative repair. The real unlock is how quickly it converges after feedback. The feedback loop speed is what changes developer throughput.

The timing of this review matters more than people realize. OpenAI's roadmap for Codex has some big shifts coming that'll make your current callouts very interesting in hindsight.