Last month, OpenAI launched GPT-5, its most advanced AI model with significant upgrades in reasoning, coding, and other key areas. At the time of launch, OpenAI claimed that GPT-5 comes with improved agentic capabilities, scoring 74.9% on SWE-bench Verified and 88% on Aider Polyglot.
Today, OpenAI announced GPT‑5-Codex, a new GPT-5 model that is optimized for agentic coding in Codex. OpenAI mentioned that this latest model was developed to perform better on real-world software engineering work, including creating new projects from scratch, adding features and tests to existing projects, performing large-scale code refactoring tasks, and more.
In SWE-bench Verified benchmarks, GPT-5 High scored 72.8% while the new GPT-5-Codex scored 74.5%. On code refactoring tasks, the latest model scored 51.3%, significantly higher than GPT-5"s 33.9%.
One of the most impressive things about GPT‑5-Codex is its ability to decide how much time to spend thinking based on the complexity of the task. During internal testing, OpenAI noticed that GPT‑5-Codex was able to work independently for more than 7 hours at a time on large, complex tasks.
GPT‑5-Codex now performs better on front-end tasks, including mobile websites. OpenAI mentioned that this model can accept images as input in the cloud, visually inspect its progress, and display screenshots of the accomplished work.
Finally, this new coding-optimized model is better at code reviews. It can match the intent of a pull request to the actual diff, reason over the entire codebase, and even execute code and tests to validate the intended behavior. OpenAI claims that it is now using Codex for code reviews internally, and it is already catching hundreds of issues every day.
Codex with the GPT‑5-Codex model is now available as part of ChatGPT Plus, Pro, Business, Edu, and Enterprise plans via the Codex CLI, IDE Extension, web, mobile, and for code reviews in GitHub. For developers using the Codex CLI via an API key, GPT‑5-Codex will be made available in the API soon.