OpenAI has unveiled GPT-5-Codex, a specialized version of GPT-5 built to solve complex software engineering problems. Designed for the Codex CLI, IDE extension, and cloud environment, this model can operate on its own for more than seven hours, providing production-ready solutions without human help. It is now the default engine in Codex’s cloud service and is accessible wherever developers use Codex.
One of the key innovations in GPT-5-Codex is adaptive reasoning. The model changes its reasoning time based on how complicated the task is:
For small, well-defined tasks, it interacts quickly in a chat-like manner.
For larger, multi-file refactors, it consistently works and allocates extra time to ensure accuracy.
Efficiency data shows that for the simplest 10% of tasks, GPT-5-Codex consumed 93.7% fewer tokens compared to GPT-5. However, for the most complex 10% of tasks, it doubled its reasoning and iteration time, leading to deeper and more accurate results.
When tested against GPT-5, GPT-5-Codex achieved 51.3% accuracy compared to GPT-5’s 33.9% on systematic, multi-step code changes. One test involved a Gitea pull request that required a context variable to pass through 232 files and over 3,500 lines of code — a task that GPT-5-Codex handled with greater accuracy and efficiency.
Beyond refactoring, GPT-5-Codex improves code review processes by:
In evaluations of recent commits from popular open-source projects, GPT-5-Codex provided more accurate and valuable review comments, cutting down on irrelevant feedback and focusing on critical issues.
GPT-5-Codex was trained using reinforcement learning on real-world coding tasks including building projects from scratch, adding features, debugging, refactoring, and generating tests. This training helps the model behave according to common coding practices, pull request conventions, and project-specific guidelines (via AGENTS.md files).
Developer Access
GPT-5-Codex is available now through:
• Codex CLI
• IDE extension
API key integration for CLI workflows is coming soon, making it easier for developers to use Codex in agentic coding scenarios.
Early Adoption Success Stories
Engineering teams are already seeing measurable benefits:
Duolingo: Aaron Wang, Senior Software Engineer, mentioned that Codex excelled in backend Python code reviews It was the only one to catch tricky backward compatibility issues and consistently found the hard bugs that other bots missed.
Cisco Meraki: A Tech Lead observed that Codex allowed their team to offload refactoring and test generation on a cross-team project. It produced high quality, fully tested code that I could quickly hand back, keeping the feature on schedule without adding risk.
With GPT-5-Codex, OpenAI is changing what AI can do in software engineering. From multi-hour autonomous coding to precise refactoring and smarter code reviews, Codex is not just an assistant — it’s becoming a <
