Openai upgrade Codex with a new version of GPT-5

Openai announced on Monday that it is releasing a new version of GPT-5 to its AI coding agent, Codex. The company says its new model, called GPT-5-CODEX, spends its “thinking” time more dynamically than the previous model, and can spend anywhere from a few seconds to seven hours on coding work. As a result, this agent performs better on coding benchmarks.

The new model is now rolling out in Codex products – which can be accessed through a terminal, IDE, zeethb, or chatgipt – all chat plus, pro, business, Edu and entry users. Openai says that it is planning to provide models to API customers in future.

The update is part of the attempt to make Openai’s codex more competitive with other AI coding products, such as cloud code, cursor of Anysphere, or Github Copilot of Microsoft. The market for AI coding tools has resulted in a lot of crowd as a result of intensive user demand in the previous year. Cursor overtook $ 500 million in the first ARR in 2025 and a similar code editor, Windsurf, was the theme of an chaotic acquisition attempt, which saw its team divided between Google and the feeling.

Openai says GPT-5-CODEX has dismissed GPT-5 Self-bench verifiedA benchmark measuring agent measures a benchmark measuring a benchmark measuring a benchmark on coding refractoring functions with coding capabilities.

The company also states that it trained the GPT-5-Codex to review the code, and asked the experience software engineers to evaluate the model review comments. Engineers allegedly found the GPT-5-Codex to present less wrong comments, while adding more “high-effects”.

In a briefing, Alexander Embiricos, head of Openai’s Codex product, said that most of the increased performance were thanks to the GPT-5-Codex’s dynamic “thinking abilities”. Users may be familiar with a router of GPT-5 in Chatgpt, which directs questions for various models based on the complexity of a task. Embiricos stated that GPT-5-Codex works equally, but there is no router below the hood, and how long can it adjust to work on a task in real time.

Embiricos states that it is an advantage compared to a router, which decides how computational power and time to use a problem to use a problem to use a problem and time. Instead, the GPT-5-CODEX can fix five minutes in a problem that needs to be spent another hour. Embiricos stated that they have seen the model above seven hours in some cases.