Vibe Check: GPT-5 Codex Can Code for 35 Minutes Straight—If You Ask Nicely
It launches today—here’s our day-zero vibe check! 🚀
What’s New
-
New Model (GPT-5 Codex): The new fine-tuned version of GPT-5 designed specifically for coding not only completes simple queries quickly but also has a knack for those multi-step refactors that formerly required your undivided attention. 🧠✨
-
Seamless Handoff: Start your coding project in VS Code and hand the rest over to the mighty Codex Cloud! Why? Because it keeps working even while your machine is taking a well-deserved nap (or you are). 😴💻
-
Better Code Reviews: OpenAI is now launching a code review bot that goes deeper than reading code alone—it's executing checks and applying fixes right on GitHub. No more lame autocorrects; we're talking about serious quality control here! 🛠️
-
Where to Find It: GPT-5 Codex will soon be the backbone of the web-based Codex. Get ready to see it in your CLI and VS Code extension, and fret not over pricing—it’s aligned with GPT-5! 🤑
Codex Learns to Think—and Keeps Going
Imagine this: Codex dynamically decides how long to think. Super helpful, right? For those trivial questions like, “What folder are we in?” it’s all about speed. However, when faced with complex requests, it can engage in deep analysis, mimicking a real-life colleague who knows when to accelerate and when to ponder. 📚🤔
But here's the kicker: by allowing you to have an interactive back-and-forth, it stands out as more than a simple command-executor. You can bounce ideas around instead of merely firing off requests.
The Handoff feature is another highlight. You can kick off a task locally in VS Code, send it to Codex Cloud, and rest assured it’ll chug along in the background, ready for you when you return. How cool is that? Think of it as handing off your tasks to a superhero while you’re away. 🦸♂️
What’s Working
🥳 Smart Thinking Time
Watch it work! Kieran Klassen, the Cora GM, found that GPT-5 Codex can intelligently balance between hustling and deliberation. When asked to explain a project, it was quick on its feet. But when challenged with a deeper task, it took the time necessary to map things out beautifully.
⏳ Long Sessions with Right Prompts
Danny Aziz, GM of Spiral, tricked it into a 35-minute work session with the right prompts. That’s progress toward persuasive autonomy! Want to know the secret? Breaking tasks into smaller milestones was the key to unlocking Codex’s long-form potential.
What Needs Work
👀 Picky Worker
Despite its new-found flexibility, Codex remains a bit persnickety. When faced with overreaching tasks, it flagged, “That’s a multi-sprint job.” 🤨 So, if your task is broad, be prepared to coax it with creative prompting!
⚙️ Environment Setup Friction
Another hiccup? Getting it to play nice with your local environment. When Kieran attempted to run Ruby on his own machine, he dove down a rabbit hole of tedious configurations just to bridge the gap between local and cloud versions. 🐇🛠️
🔄 Multi-Agent Workflows
Kieran encountered trouble running multiple review agents, as Codex paused after each task for permission rather than continuing autonomously. This hitch can frustrate those aiming for a fluid workflow, especially those accustomed to Claude Code’s competence in maintaining pace. 🚦
The Final Vibe Check
While GPT-5 Codex shows promise with dynamic thinking and effortless cloud transitions, it’s clear it hasn’t fully evolved just yet. It's paving the way for a future where AI not only assists with tasks but truly collaborates. Just stay on your toes—it might still need a nudge or two to reach its full potential! ✨
In conclusion, it’s becoming an invaluable asset to our engineering toolkit, but we’re still waiting on a few friendly improvements.
Written by Dan Shipper
Dan Shipper is the CEO and cofounder of Every. Catch his insights on Twitter and tune into the AI & I podcast! 🎙️