OpenAI's GPT-5 Codex can code autonomously for 7 hours straight

GPT-5 Codex breaks all records: 7 hours of autonomous coding, 15x faster on simple tasks, 102% more thinking on complex problems. OpenAI engineers now refuse to work without it.

GPT-5 Codex shatters records with 7-hour autonomous coding sessions, dynamic thinking that adjusts effort in real-time, and code review capabilities that caught OpenAI's own engineers off guard.

The coding agent revolution just hit hyperdrive. OpenAI released GPT-5 Codex yesterday, and Sam Altman wasn't exaggerating when he tweeted the team had been "absolutely cooking." This isn't just another incremental update—it's a fundamental shift in how AI approaches software development, with the model working autonomously for up to 7 hours on complex tasks.

The 7-hour coding marathon

Just weeks ago, Replit set the record with Agent 3 managing 200 minutes of continuous independent coding. GPT-5 Codex just obliterated that benchmark, working for 420 minutes straight.

OpenAI team members revealed in their announcement podcast: "We've seen it work internally up to 7 hours for very complex refactorings. We haven't seen other models do that before."

The numbers tell a shocking story. While standard GPT-5 uses a model router that decides computational power upfront, Codex implements dynamic thinking—adjusting its reasoning effort in real-time. Easy responses are now 15 times faster. For hard problems, Codex thinks 102% more than standard GPT-5. Developer Swyx called this "the most important chart" from the release: "Same model, same paradigm, but bending the curve to fit the nonlinearity of coding problems."

The benchmarks barely capture the improvement. While Codex jumped modestly from 72.8% to 74.5% on SWE-bench Verified, OpenAI's custom refactoring eval shows the real leap: from 33.9% to 51.3%.

Early access developers are losing their minds. Nick Doobos writes it "hums away looking through your codebase, and then one-shots it versus other models that prefer immediately making a change, making a mess, and then iterating." Michael Wall built things in hours he never thought possible: "Lightning fast natural language coding capabilities, produces functional code on the first attempt. Even when not perfectly matching intent, code remains executable rather than broken." Dan Shipper's team ran it autonomously for 35 minutes on production code, calling it "a legitimate alternative to Claude Code" and "a really good upgrade."

Why it thinks like a developer

GPT-5 Codex doesn't just code longer—it codes smarter. AI engineer Daniel Mack calls this "a spark of metacognition"—AI beginning to think about its own thinking process.

The secret weapon? Code review capabilities that OpenAI's own engineers now can't live without. Greg Brockman explained: "It's able to go layers deep, look at the dependencies, and raise things that some of our best reviewers wouldn't have been able to find unless they were spending hours." When OpenAI tested this internally, engineers became upset when it broke. They felt like they were "losing that safety net." It accelerated teams, including the Codex team itself, tremendously. This solves vibe coding's biggest problem. Andre Karpathy coined the term in February: "You fully give into the vibes, embrace exponentials, and forget that the code even exists. When I get error messages, I just copy paste them in with no comment."

Critics said vibe coding just shifted work from writing code to fixing AI's mistakes. But if Codex can both write and review code at expert level, that criticism evaporates.

The efficiency gains are unprecedented. Theo observes: "GPT-5 Codex is, as far as I know, the first time a lab has bragged about using fewer tokens." Why spend $200 on a chunky plan when you can get the same results for $20? Usage is already up 10x in two weeks according to Altman. Despite Twitter bubble discussions about Claude, a PhD student named Zeon reminded everyone: "Claude is minuscule compared to Codex" in real-world usage.

The uneven AI revolution

Here's the uncomfortable truth: AI's takeoff is wildly uneven. Coders are living in 2030 while everyone else is stuck with generic chatbots.

Professor Ethan Molick doesn't mince words: "The AI labs are run by coders who think code is the most vital thing in the world... every other form of work is stuck with generic chat bots."

Roon from OpenAI countered that autonomous coding creates "the beginning of a takeoff that encompasses all those other things." But he also identified something profound: "Right now is the time where the takeoff looks the most rapid to insiders (we don't program anymore, we just yell at Codex agents) but may look slow to everyone else."

This explains everything. While pundits debate AI walls and plateaus, developers are experiencing exponential productivity gains. Anthropic rocketed from $1 billion to $5 billion ARR between January and summer, largely from coding. Bolt hit $20 million ARR in two months. Lovable and Replit are exploding. The market has spoken. OpenAI highlighted coding first in GPT-5's release, ahead of creative writing. They're betting 700 million new people are about to become coders.

Varun Mohan sees the future clearly:

"We may be watching the early shape of true autonomous dev agents emerging. What happens when this stretches to days or weeks?"

The implications transcend coding. If AI can maintain focus for 7 hours, adjusting its thinking dynamically, we're seeing genuine AI persistence—not just intelligence, but determination. The gap between builders and everyone else has never been wider. But paradoxically, thanks to tools like Lovable, Claude Code, Cursor, Bolt, and Replit, the barrier to entry has never been lower.

The coding agent revolution isn't coming. For those paying attention, it's already here.