OPENAI UPDALS CODEX with a fresh version of GPT-5

Opeli announced on Monday that he was transferring a fresh version of GPT-5 to AI, Codex coding agent. The company claims that its fresh model, called GPT-5-Codex, spends its “thinking” more dynamically than previous models and can spend from a few seconds to seven hours on coding. As a result, it works better in the field of comparative agency coding.

The fresh model now implements in Codex products – to which you can access through the terminal, IDE, GitHub or ChatgPT – to all chatgpt plus, pro, business, edu and enterprise users. Opeli claims that he plans to share the model with API clients in the future.

The update is part of the OpenAI efforts to make Codex more competitive with other products encoding artificial intelligence, such as Claude Code, Anysphere cursor or Microsoft Github Copilot. The market of artificial intelligence coding tools has become much more crowded over the past year, as a result of intensive user demand. The cursor exceeded $ 500 million in ARR earlier in 2025, and Windsurf, a similar code editor, was the subject of a messy acquisition test in which his team shared between Google and Poznań.

Opeli says that GPT-5-Codex exceeds GPT-5 Swo-Bench verifiedBenchmark measurement Agentic Coding coding capacity, as well as reference efficiency of measuring the tasks of re -invoicing code from gigantic, fixed repositories.

The company also claims that it has trained GPT-5-Codex in conducting code reviews and asked software engineers to evaluate model review comments. Engineers reportedly found GPT-5-Codex to send less incorrect comments while adding “high-influence comments”.

On the briefing, the main product Codex Openai Alexander Embiricos said that a significant part of increased performance was thanks to the animated “thinking ability” GPT-5-Codex. Users can know the GPT-5 router in ChatgPT, which directs queries to various models depending on the complexity of the task. Embiricos said that GPT-5-Codex works similarly, but there is no router under the hood and can adapt to working on real-time task.

Embiricos claims that this is an advantage compared to a router that determines how much computing force and time to utilize at the beginning. Instead, GPT-5-Codex can decide for five minutes for a problem that he has to spend another hour. Embiricos said that in some cases he saw the model took over seven hours.