The news of September 16 is thatOpenAI This morning we announced a new launch GPT-5-Codex♪ And already in their AI Programming Assistant Codex centerline. This is a new version of GPT-5 that will be optimized for Agent programming。
Currently GPT-5-Codex is open through Codex to ChatGPT Plus, Pro, Business, Edu and Enterprise users and can be used at terminals, IDE, GitHub or ChatGPT. 1AI has been officially informed that OpenAI also plans to open the model to API clients in the future。
OpenAI states that a major improvement in GPT-5-Codex is in dynamic computing capability. The model allows for a more flexible allocation of “thinking time”, which may range from seconds to up to seven hours in the performance of a code assignment, thus showing better performance in multiple programming baseline tests。
GPT-5-Codex was able to adjust the time and effort required in real time in the course of a mission, unlike the way in which it had previously relied on a “routing” mechanism to predict the complexity of the mission and allocate resources. For example, it may take up to an additional hour, or even up to seven hours in individual cases, to determine five minutes after the start. This mechanism has made it more advantageous in complex programming tasks。
According to OpenAI data, GPT-5-Codex is better than the standard GPT-5 model in the SWE-bench Verified test and in the remodeling of large code libraries。
In addition, the model was specially trained in code review and assessed by senior software engineers. The results show that GPT-5-Codex has fewer errors, while providing more “high-impact comments”。
In his briefing, Alexander Embiricos, Director of OpenAI Codex Products, stated: “GPT-5-Codex's ability to think dynamically has increased significantly. It does not depend on a fixed route, but is able to judge the need for additional working hours in the middle of a mission, which is particularly advantageous on complex issues.”