OpenAI's New GPT-5-Codex Can Spend Hours Solving Complex Coding Tasks

OpenAI has introduced a new version of its agentic coding model called GPT-5-Codex, which is now being added to Codex products across different platforms. The model is available through the terminal, IDEs, GitHub, and ChatGPT for Plus, Pro, Business, Edu, and Enterprise users. OpenAI says the model will also be made available to API customers at a later stage.

GPT-5-Codex is designed to adjust how long it spends working on a coding task in real time, ranging from just a few seconds to seven hours. OpenAI claims the model performs better than GPT-5 on benchmarks like SWE-bench Verified, which track real-world coding abilities. Engineers found that the model made fewer wrong comments and provided more valuable feedback than earlier versions.

"GPT‑5-Codex adapts how much time it spends thinking more dynamically based on the complexity of the task," OpenAI noted. "The model combines two essential skills for a coding agent: pairing with developers in interactive sessions, and persistent, independent execution on longer tasks."

According to Alexander Embiricos, product lead at OpenAI's Codex, this improvement is largely due to GPT-5-Codex's dynamic "thinking" process. He explained that GPT-5 uses a system to decide at the beginning of a task how much time and computing power to use; in contrast, GPT-5-Codex can change its resources in real time. For example, if the model realizes after five minutes that it needs more time, it can extend its reasoning and keep working, sometimes for hours, until it finds a better solution.