On Monday, OpenAI announced the release of a new version of GPT-5 tailored for its AI coding agent. The company revealed that the new model, named GPT-5-Codex, offers greater flexibility in processing time compared to previous versions, with code execution times ranging from seconds to up to seven hours. This enhanced adaptability enables superior performance in agent-based coding benchmarks.

The updated model is now available within the Codex product suite, accessible via terminal, IDE, GitHub, or ChatGPT. It is available to all ChatGPT Plus, Pro, Business, Edu, and Enterprise users. OpenAI has also indicated plans to extend access to API customers in the future.

This release is part of OpenAI’s broader strategy to strengthen Codex’s competitiveness against other AI coding tools such as Claude Code, Cursor from Anysphere, and GitHub Copilot by Microsoft. The AI coding tools market has grown increasingly crowded due to strong user demand. For instance, Cursor achieved over $500 million in annual recurring revenue by early 2025, while Windsurf, another coding editor, became the center of a contentious acquisition attempt that ultimately saw its team split between Google and Cognition.

According to OpenAI, GPT-5-Codex outperforms GPT-5 on key benchmarks including SWE-bench Verified, which measures agent coding capabilities, and a benchmark focused on refactoring tasks in large, mature codebases.

In addition, OpenAI noted that GPT-5-Codex has been trained to conduct code reviews and invited experienced software engineers to evaluate the quality of its review comments. Engineers reportedly found that GPT-5-Codex produced fewer erroneous comments while generating more “high-impact comments.”

Durring a briefing, Alexander Embiricos, OpenAI’s product lead for Codex, explained that much of the performance gain stems from GPT-5-Codex’s dynamic “thinking capability.” Users may already be familiar with the GPT-5 router in ChatGPT, which directs queries to different models based on task complexity. According to Embiricos, GPT-5-Codex operates in a similar fashion but without an internal router, and it can adjust the time allocated to a task in real time.

He noted that this represents a key advantage over traditional routers, which determine the amount of computational resources and time to apply at the outset. In contrast, GPT-5-Codex can, for example, decide after five minutes into a task that it needs an additional hour to complete. Embiricos mentioned that he has observed the model spending more than seven hours on certain complex tasks.