NEW GPT-5 Codex Model (Tested for 6 Hours)


Hey Reader,

Last night, OpenAI dropped something that made me reconsider everything.

A special variant of GPT-5 called "Codex", purpose-made for AI coding.

And after testing it non-stop 6+ hours, I need to tell you something important:

This is an actual game-changer. It's the real deal.

Here's What Just Happened

While everyone was sleeping, OpenAI quietly released three new models

  • gpt-5-codex-high
  • gpt-5-codex-medium
  • gpt-5-codex-low

But here's the insane part:

I've been hammering the "High" version for 6+ hours straight on ChatGPT Plus.

Building apps. Breaking things. Pushing limits.

... and I still haven't hit a single usage limit.

Not. One. Time.

Claude Is... Out?

Remember when I said Claude Opus 4 was the best AI coder?

Or that you needed a $200/month Cursor subscription?

Yeah, about that...

GPT-5 Codex, at least today, has been unlimited for me.

And it's included in your ChatGPT Plus subscription for $20...

Let me repeat that: Twenty. Dollars. 😂

But There's a Catch (Of Course)

Right now, it only works in two places:

1️⃣ Codex CLI - The terminal thing that scares most people

2️⃣ Codex IDE Extension - Works with VS Code and Cursor

The model does not exist in any other coders yet.

My AI Coding Blueprint now includes Weekly Q&A Live Sessions (every Thursday, 11 AM ET) where you can jump on, submit your questions, and I'll answer them live on Zoom. I can help you debug your app, understanding a new service, or fix a problem. And the price? It's still just $997 one-time for the Group Mentorship plan!

What's New In GPT-5 Codex?

GPT-5 Codex appears to have a much larger token context window (some say 1 million, unconfirmed).

The model also thinks... dynamically.

Sometimes it answers in 5 seconds, other times it disappears for ages

I watched it think for 10, 15 and 23 minutes today, then come back with a complete features without problems.

It's like having a junior developer who randomly takes coffee breaks but comes back with perfect code.

Do This Right Now

If you have ChatGPT Plus (which you should), here's your move:

  1. Install the Codex extension or Codex CLI (it's free)
  2. Start with something simple to test it
  3. Watch it blow your mind

I'm not exaggerating.

This thing has instantly become my default coding model.

And if you ask me... Claude 4.5 (expected this week) is not going to fix this.

The Bottom Line

This is the one. The model to pick for the foreseeable future.

I'm working on a full breakdown video.

But I couldn't wait to tell you.

Talk soon,
Rob

P.S. My AI Coding Blueprint students watched me test GPT-5 Codex live all day today. They knew all of this hours ago, and they're already making moves. If you've been waiting for the right time to join, this is it. The tools just got 10x better, and the price is still the same. For now.

Robin Ebers

Coder of 20+ years teaching non-technical people how to build their own software business in 30 days with AI. No devs or code required.

Read more from Robin Ebers

Hey Reader I need to confess something embarrassing. For years, I was the worst kind of engineer. The kind who spent 100 hours perfecting code for apps that never launched. Clean architecture. Zero bugs. Fully scalable. And completely worthless, because nobody used them. Here's The Painful Truth While I was busy "doing it right"... Non-technical founders were shipping garbage code and making money. My perfect projects were sitting in GitHub repos. Their messy apps took daily payments. The...

Hey Reader, I just spent hours building the exact same app twice. Same prompt. Same tech stack. Same features. The only difference was that one used Claude Sonnet 4.5 (Claude Code), and the other used GPT-5 (Codex CLI). And holy sh*t, the results were not even close. Let me show you what happened. The Test I built a simple to-do list app with user authentication. Nothing crazy. Just login, add tasks, edit them, delete them. I even told both models: "Keep it super secure." Should be easy...

Hey Reader, Two days ago my student messaged me. He was almost angry... "GPT-5 worked for 23 minutes and then said it doesn't understand!" Oof. That's rough, and I know how this feels. But that's when it hit me. We're all prompting AI wrong. Everyone of us. The $200,000 Mistake Big tech pays prompt engineers $200k+/year. Because they know what a difference of good prompts. Imagine telling your GPS "go somewhere nice" instead of an actual address One gets you lost. One gets you there. So I did...