OpenAI has unveiled a major upgrade to its AI coding agent, Codex, with the release of GPT-5-Codex. The new model is designed to “think” more dynamically than previous versions, allowing it to spend anywhere from a few seconds to seven hours tackling a coding task. According to the company, this improvement makes the AI significantly more effective on agentic coding benchmarks, giving developers a powerful tool for complex programming challenges.
GPT-5-Codex is now available in all Codex products, which can be accessed through terminals, integrated development environments (IDEs), GitHub, or directly via ChatGPT. The rollout includes ChatGPT Plus, Pro, Business, Edu, and Enterprise users, with plans to extend access to API customers in the near future. OpenAI says this move aligns with its goal of expanding Codex’s reach and usability among professional developers.
The update positions Codex to compete more aggressively in the growing AI coding market, which now includes tools like Claude Code, Anysphere’s Cursor, and Microsoft’s GitHub Copilot. Market demand has surged in the past year, with Cursor reaching $500 million in annual recurring revenue and Windsurf experiencing a high-profile acquisition that split its team between Google and Cognition. GPT-5-Codex is OpenAI’s answer to this crowded landscape, offering advanced capabilities for professional coding workflows.
Performance improvements are particularly notable in code refactoring and code review tasks. OpenAI trained GPT-5-Codex to conduct code reviews, and experienced software engineers reported that the AI produced fewer incorrect comments while generating more “high-impact” feedback. This suggests the model could save teams time and improve overall code quality.
Alexander Embiricos, OpenAI’s Codex product lead, explained that the model’s dynamic “thinking abilities” are key to its success. Unlike GPT-5 in ChatGPT, which uses a router to assign computational resources upfront, GPT-5-Codex can adapt in real time, deciding during a task whether to dedicate more time and processing power. Embiricos noted that the model has taken as long as seven hours on particularly complex problems, demonstrating an unprecedented level of flexibility for AI coding agents.
source: techcrunch