OpenAI rolls out GPT-5 codex for improved coding

OpenAI rolled out something new for the coding crowd: GPT-5-Codex, a souped-up version of its AI coding agent that’s designed to think harder, longer, and, if necessary, seven hours straight. 

Yes, seven. Imagine your rubber duck debugger, but one that sometimes disappears into the basement to brood over your problem until it re-emerges with a solution.

The upgrade brings a major change: instead of allocating a fixed “brainpower budget” at the start of a task, GPT-5-Codex can adjust its effort on the fly. 

Maybe it spends just a few seconds fixing a typo, or maybe it camps out for hours wrangling a gnarly refactor. 

According to OpenAI, this flexibility helps the model ace tricky agentic coding benchmarks and even makes it less annoying during code reviews, where human engineers said it now adds more helpful comments and fewer “um, actually” moments.

Starting this week, GPT-5-Codex is landing in all the usual OpenAI haunts, your terminal, IDE, GitHub, and ChatGPT, for Plus, Pro, Business, Edu, and Enterprise users. API access is on the roadmap, though no date is set.

The timing is no accident. AI coding assistants have gone from shiny novelties to big business, with competition heating up fast. 

Cursor, a code editor from Anysphere, cracked half a billion in annual revenue earlier this year, while Windsurf, another buzzy editor, sparked a messy tug-of-war between Google and Cognition. 

Add in rivals like Claude Code and Microsoft’s GitHub Copilot, and Codex is suddenly in a knife fight for developer mindshare.

Alexander Embiricos, OpenAI’s Codex lead, framed the new model’s flexibility as its secret weapon. 

Unlike GPT-5’s router system, GPT-5-Codex simply decides mid-task whether it should keep grinding. 

That means it can shift gears when it realizes your “simple bug fix” is secretly a cursed spaghetti monster from 2013.

So yes, Codex might now spend hours chewing over your code. 

But if the result is cleaner pull requests and fewer facepalms during review, developers may just forgive it for occasionally working harder than they do.

Will GPT-5-Codex’s ability to spend hours on complex coding problems make it a game-changer for developers, or does this just mean longer wait times for solutions that human programmers could handle more efficiently? Should we be concerned about AI coding assistants becoming so sophisticated that they reduce opportunities for junior developers to learn through hands-on problem-solving? Tell us below in the comments, or reach us via our Twitter or Facebook.