OpenAI has announced an update to its Codex programming tool with a new version of GPT-5. The company says the new model, called GPT-5-Codex, is able to use its “think time” more efficiently, spending anywhere from a few seconds to seven hours on a task. This has significantly improved its performance in agent coding tests.
OpenAI has announced an update to its Codex programming tool with a new version of GPT-5. The company says the new model, called GPT-5-Codex, is able to use its “think time” more efficiently, spending anywhere from a few seconds to seven hours on a task. This has significantly improved its performance in agent coding tests.
The new model is rolling out now. It’s available in Codex products — which can be accessed via the terminal, IDE, GitHub, or ChatGPT — for all ChatGPT Plus, Pro, Business, Edu, and Enterprise users. OpenAI also announced plans to make the model available to API clients in the future, TechCrunch reports .
According to OpenAI, the GPT-5-Codex model outperformed GPT-5 in the SWE-bench Verified test, which measures how well an AI handles agent-based coding. It also performed better in code refactoring tasks from large repositories.
The company also said it trained GPT-5-Codex to conduct code reviews and then asked experienced software engineers to evaluate the model's comments. They noted that GPT-5-Codex provided fewer erroneous comments while offering more "high-quality comments" that had significant impact.
OpenAI’s Codex lead, Alexander Embiricos, said that the significant performance gains in GPT-5-Codex are due to its dynamic “thinking ability.” Users may be familiar with the “router” in GPT-5 for ChatGPT, which routes requests to different models based on the complexity of the task. Embiricos explained that GPT-5-Codex works similarly, but it doesn’t have a built-in router and can adjust the time it takes to complete a task in real time.
Embiricos says this is an advantage over a router, which determines the amount of processing power and time to solve a problem at the outset. Instead, GPT-5-Codex might decide after five minutes of work on a task that it needs another hour. Embiricos says he has seen the model work for more than seven hours in some cases.
Recall that OpenAI launched Codex in May of this year. At that time, the company announced that the AI agent works best with Python, and also supports over a dozen other programming languages, including JavaScript, Go, Perl, PHP, Ruby, Swift, Shell, and others.
Comment hidden for violating commenting rules.