OpenAI upgrades Codex with a new version of GPT-5

OPNA announced Monday that it was releasing a new version of GPT -5 on its AI Coding Agent, CodesThe The company says that the new model known as GPT -5 -Codex spends its “thoughtful” time dynamically compared to previous models and can spend on a coding task for a few seconds to seven hours. As a result, it performs better in agent coding benches.

The new model is now rotating on codes products – all ChatzPT Plus, Pro, Business, EDU and Enterprise users can be accessed through a terminal, IDE, githabp or chatzPT. Open says that API customers have plans to make the model available in the future.

The update is the update to make the Codex more competitive with other AI coding products such as Clod Code, Carser of any Spier or Githab Copilot in Microsoft. Due to the intense user demand, the market for AI coding equipment has become much more crowded last year. Cursor Exceeded $ 500 million from ARR Before that, in 2025, and similar code editor was subject to Windsorf An attempt to acquire a chaos It saw its team divided between Google and Knowledge.

OpenAI says that GPT -5 -Codex exceeds GPT -5 Has been verified in the needle-benchA benchmark measuring agent is the coding capacity, as well as the performance of a criteria in the refactoring functions of the codes from the large, established storage.

The company also said that it trained GPT -5 -Codex to manage the code review and asked to evaluate the models of the experience software engineers. Engineers are reported to have found GPT-5-Codex for submitting less wrong comments while adding more “high-influence comments”.

In a briefing, Alexander Embirikos, the Codex Product Leader of OpenAIA, said that most of the extended performance was thanks to the dynamic “thinking skills of the GPT -5 -Codex”. Users can be familiar with the GPT -5 router in ChatGPT, which indicates questions in various models based on the complexity of a job. Embiricos says that GPT -5 -Codex works the same way, but there is no router under the hood and can adjust how long it should work in a real time.

Embiricos says it is an advantage compared to a router, which determines how much calculating power and time to use for a problem in the beginning. Instead, GPT -5 -Codex can decide a problem for five minutes that will spend another hour to spend it. Embiricos said he had seen the model taking for seven hours in some cases.