OpenAI has released GPT-5-Codex, a version of its popular model optimized for complex software engineering tasks. The new model can operate autonomously for over seven hours to deliver working solutions without human intervention. It’s now the default in Codex’s cloud service and available on various platforms.
GPT-5-Codex boasts adaptive reasoning, which adjusts its processing time based on task complexity. This allows it to tackle both small, well-defined tasks and large-scale refactoring work with ease. The model has been shown to achieve higher accuracy than GPT-5 in direct evaluations of refactoring tasks, with an accuracy rate of 51.3% compared to 33.9%.
The model’s efficiency gains have also been impressive, using 93.7% fewer tokens for simple requests and allocating more effort for complex ones. This makes it a valuable tool for developers looking to streamline their workflow.
GPT-5-Codex has also been trained to strengthen code review workflows, navigating repositories, analyzing dependencies, and running tests to validate correctness. It’s reported that the model produces high-quality review comments that are accurate and higher-value, reducing noise for developers and highlighting critical issues.
The model was trained using reinforcement learning on real-world coding tasks, aligning its behavior with common coding styles and pull request conventions. It can also follow project-specific guidelines defined in AGENTS.md files.
Developers can access GPT-5-Codex through the Codex CLI and IDE extension, with API key access expected in upcoming releases. Early users have praised the model’s performance, with one engineer noting that it caught tricky backward compatibility issues and consistently found hard bugs that other bots missed.
Source: https://www.infoq.com/news/2025/09/gpt-5-codex