OpenAI's GPT-5-Codex Boosts Efficiency with Dynamic Thinking Time

OpenAI has announced the release of a new version of its AI programming tool, Codex, named GPT-5-Codex. This new model features a more dynamic allocation of "thinking" time, allowing it to complete programming tasks in a range of a few seconds to 7 hours. This dynamic capability enhances its performance in agentic coding benchmarks, making it a more efficient tool for developers.

GPT-5-Codex is now being gradually rolled out across the Codex product line. Users can access the tool through terminals, integrated development environments (IDEs), GitHub, or ChatGPT. All ChatGPT Plus, Pro, Business, Edu, and Enterprise users have access to this new model, with plans to extend its availability to API clients in the future.

This update is part of OpenAI's strategy to enhance the competitiveness of Codex against other AI programming products such as Claude Code, Anysphere's Cursor, and Microsoft GitHub Copilot. The AI programming tool market has seen intense competition, driven by surging user demand. The new model has shown superior performance in two key tests: the SWE-bench Verified benchmark, which measures agentic programming capabilities, and a performance test for code refactoring tasks in large, mature codebases.

GPT-5-Codex has undergone specialized training for code review tasks. OpenAI has invited senior software engineers to evaluate the model's review suggestions, finding that it provides fewer erroneous opinions and more "high-impact" suggestions that significantly improve code quality, security, and efficiency.

During a briefing, the Codex product lead highlighted that the performance improvements in GPT-5-Codex are largely due to its dynamic "thinking" capability. Users familiar with the routing mechanism in ChatGPT's GPT-5 will recognize a similar concept, where queries are allocated to different models based on task complexity. However, GPT-5-Codex operates without an internal routing mechanism, instead adjusting the time spent on a task in real-time.

This dynamic adjustment is a significant advantage over traditional routing mechanisms, which determine the required computational resources and time at the task's outset. GPT-5-Codex can reassess after 5 minutes of processing and decide whether to invest an additional hour to complete the task. In some cases, the model has taken over 7 hours to process a single task, demonstrating its flexibility and depth in handling complex programming challenges.