OpenAI releases new GPT-5-Codex model for autonomous coding and large-scale refactoring

OpenAI has released GPT-5-Codex, an upgraded version of GPT-5 built specifically for agentic coding in Codex. The model is designed to handle real-world engineering work, from starting new projects and adding features to writing tests and performing large-scale refactoring. It outperforms earlier GPT-5 variants on key benchmarks, including a 74.5% score on SWE-bench Verified compared to 72.8% for GPT-5 High, and 51.3% on refactoring tasks compared to 33.9%. It also improves front-end development, including mobile web support.

GPT-5-Codex introduces stronger autonomy, able to run on complex tasks for more than seven hours while adjusting how much time it spends based on task complexity. In the cloud, it can take images as input, track progress visually, and generate screenshots of completed work. These capabilities extend Codex beyond quick interactions into sustained, independent development support.

The model is also better at code reviews by aligning pull request intent with diffs, reasoning across entire repositories, and executing code and tests to validate results. The model is now included with ChatGPT Plus, Pro, Business, Edu, and Enterprise plans across CLI, IDE, web, mobile, and GitHub, with API access for Codex CLI coming soon.