GPT-5-Codex, the new artificial intelligence (AI) model from OpenAI, can write computer code for over seven hours, fix its own mistakes, and complete projects without asking programmers for guidance.
The newly improved model, released on September 15, outperforms its predecessor on key coding benchmarks. GPT-5-Codex achieves 74.5% accuracy on SWE-bench Verified compared to GPT-5’s 72.8%. Code refactoring tasks show even bigger gains, jumping from 33.9% to 51.3% success rates.
Source: OpenAI
“During testing, we’ve seen GPT-5-Codex work independently for more than 7 hours at a time on large, complex tasks, iterating on its implementation, fixing test failures, and ultimately delivering a successful implementation,” OpenAI stated in its announcement.
The model uses what OpenAI refers to as “adaptive thinking.” Simple coding questions get quick responses with about 93.7% fewer processing tokens than GPT-5. Complex projects trigger extended reasoning sessions that can last twice as long as the standard model.
Alexander Embiricos, OpenAI’s Codex product lead, explained that GPT-5-Codex can change course mid-task. “GPT-5-Codex shines in thinking hard on complex tasks, and then quickly going back and forths with me to make tweaks or answers questions—which I ask all the time in Rust,” he wrote on various social media platforms.
Companies are already seeing results. Aaron Wang, senior software engineer at Duolingo, said GPT-5-Codex “was the only one to catch tricky backward compatibility issues and consistently found the hard bugs that other bots missed” in backend Python code reviews.
Virgin Atlantic’s VP Data & AI Richard Masters reported that “Codex isn’t only writing code, it’s actively shaping solutions alongside the team, from early design & communication through to implementation.”
At Cisco Meraki, Tech Lead Tres Wong-Godfrey used the tool for cross-team code updates. “With Codex, I offloaded the refactoring and test generation while focusing on other priorities. It produced high-quality, fully tested code that I could quickly hand back — keeping the feature on schedule without adding risk.”
The model works across multiple environments, including command line interfaces, code editors, web platforms, and GitHub. GPT-5-Codex is available immediately through ChatGPT Plus, Pro, Business, Education, and Enterprise plans. OpenAI plans to add API access later.
OpenAI reports that Codex usage grew tenfold in the past month among developers. The company positions the tool as a coding partner rather than a replacement, running in sandboxed environments with network access disabled by default for security.