OpenAI upgrades Codex with a new version of GPT-5

On Monday, OpenAI unveiled a significant upgrade to its AI coding tool, Codex, with the launch of the GPT-5-Codex model. This latest iteration boasts a more dynamic approach to processing coding tasks, allowing it to allocate between a few seconds to an impressive seven hours on a single assignment. This flexibility has led to improved performance on various coding benchmarks that assess agentic capabilities. The rollout of GPT-5-Codex is now accessible through multiple Codex products, including terminal interfaces, Integrated Development Environments (IDEs), GitHub, and ChatGPT, benefiting users across ChatGPT Plus, Pro, Business, Edu, and Enterprise tiers. OpenAI has indicated plans to extend access to API customers in the near future. This update is part of OpenAI's strategy to enhance Codex's competitiveness in a rapidly expanding market of AI coding tools. Competitors like Claude Code, Anysphere’s Cursor, and Microsoft’s GitHub Copilot have intensified the landscape, following significant user demand. Earlier in 2025, Cursor reported an annual recurring revenue exceeding $500 million, while Windsurf faced a tumultuous acquisition scenario that split its team between Google and Cognition. Notably, GPT-5-Codex has demonstrated superior performance compared to its predecessor, GPT-5, on the SWE-bench Verified benchmark, which evaluates agentic coding skills, and on tasks related to code refactoring from established codebases. Additionally, OpenAI has trained GPT-5-Codex to perform code reviews. Feedback from experienced software engineers revealed that the model produced fewer incorrect comments and contributed more impactful insights. In a recent briefing, Alexander Embiricos, the product lead for Codex at OpenAI, highlighted the model's advanced 'thinking abilities' as a key factor in its enhanced performance. Unlike the existing routing system in ChatGPT—which directs queries based on task complexity—GPT-5-Codex operates without a router, dynamically adjusting the time and resources it allocates to a problem as it progresses. Embiricos noted instances where the model has extended its work on tasks for over seven hours, showcasing its adaptability and efficiency in tackling complex coding challenges.