OpenAI begins rollout of GPT5Codex for long-running coding tasks

OpenAI on Monday began rolling out GPT5Codex, a new version of its GPT5 model optimized for automated, long-running software engineering tasks, to its Codex products across ChatGPT paid plans and developer tools. The company said the model, which is available through Codex in terminals, integrated development environment extensions, web and mobile interfaces, and GitHub, is designed to spend variable amounts of time reasoning on code problems — from seconds for small requests to many hours for large refactors — and that API access will follow.

OpenAI said GPT5Codex outperforms prior GPT5 variants on several agentic coding benchmarks and real-world engineering work. The company reported a 74.5 score on SWEbench Verified for GPT5Codex and a marked improvement on an internal refactoring benchmark, where the model scored 51.3 versus 33.9 for GPT5. OpenAI also said engineers testing the model found it produces fewer incorrect code-review comments while providing more high-impact suggestions, and that it is already catching hundreds of issues per day in internal use.

A key technical difference, OpenAI’s Codex product lead Alexander Embiricos said, is that GPT5Codex can dynamically...