OpenAI has announced GPT-5-Codex, its new model optimized for software developers. The model offers significant advantages over the previous version in code writing, editing, and reviewing. This new version is a new version of OpenAI’s coding tool, Codex. The company stated that it developed the model to perform better in real-world software engineering tasks.
GPT-5-Codex unveiled
The new model offers significant advantages in creating projects from scratch and adding features and tests to existing projects. In tests, GPT-5-Codex achieved a SWE-bench Verified score of 74.5, surpassing GPT-5 High’s score of 72.8.
The performance difference was even more pronounced in code restructuring tasks. The new model achieved a score of 51.3% in this area, while GPT-5 performed only 33.9%.
A notable feature of GPT-5-Codex is its ability to autonomously decide how long to think based on the complexity of the task. In internal tests, the model performed autonomously for over seven hours on large, complex tasks.
The model also offers improvements in front-end tasks, particularly on mobile websites. Users can now provide visual input to the model via the cloud, visually review progress, and take screenshots of completed work.
The new model also excels in code reviews. It matches the intent of pull requests with actual diffs, analyzes the entire codebase, and runs and tests the code to verify the correctness of the desired behavior. OpenAI has begun using this model for internal code reviews and reported detecting hundreds of bugs per day.
The new model is available for code reviews on web, mobile, and GitHub, as part of the ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. API access for developers using the Codex CLI will be available soon.
