OpenAI announced on Monday the launch of a new GPT-5 version for its AI programming tool Codex. The company stated that this new model, called "GPT-5-Codex," features more dynamic allocation of "thinking" time compared to previous models, with programming task completion times ranging from several seconds to up to 7 hours. This characteristic enables superior performance in agentic coding benchmarks.
Currently, GPT-5-Codex is being gradually rolled out across the Codex product line. Users can access the tool through terminals, integrated development environments (IDEs), GitHub, or ChatGPT, with availability for all ChatGPT Plus, Pro, Business, Edu, and Enterprise users. OpenAI stated that it plans to make the model available to API customers in the future.
This update represents one of OpenAI's initiatives to enhance Codex's competitiveness, aiming to compete with other AI programming products such as Claude Code, Anysphere's Cursor, and Microsoft's GitHub Copilot. Over the past year, driven by surging user demand, competition in the AI programming tools market has intensified significantly: In early 2025, Cursor's annual recurring revenue (ARR) exceeded $500 million, while another similar code editor, Windsurf, became the focus of multiple acquisition attempts, with its team ultimately splitting and joining Google and Cognition respectively, resulting in a chaotic acquisition process.
OpenAI reported that GPT-5-Codex outperformed standard GPT-5 in two key tests: the SWE-bench Verified benchmark measuring agentic programming capabilities, and performance tests for code refactoring tasks in large, mature codebases.
The company also revealed that GPT-5-Codex underwent specialized training for code review tasks, with OpenAI inviting senior software engineers to evaluate the model's review feedback. Engineers found that GPT-5-Codex provided fewer incorrect opinions while offering more "high-impact feedback" (suggestions that significantly improve code quality, security, or efficiency).
During a briefing, Alexander Embiricos, OpenAI's Codex product lead, stated that GPT-5-Codex's performance improvements largely stem from its dynamic "thinking capability." Users may be familiar with GPT-5's "routing mechanism" in ChatGPT, which allocates queries to different models based on task complexity. Embiricos noted that GPT-5-Codex operates similarly but without internal routing mechanisms, instead dynamically adjusting the time spent processing tasks in real-time.
Embiricos described this feature as a significant advantage over routing mechanisms: while routing mechanisms determine the computational resources and time needed at the task's initial stage, GPT-5-Codex can autonomously decide after 5 minutes of processing whether to invest another hour to complete the task. He mentioned that in some cases, the model spent more than 7 hours processing a single task.