OpenAI upgrades Codex with a new version of GPT-5 | TechCrunch

Openai announced on Monday that it will release a new version of GPT-5 to its AI encoding agent code. The company says its new model, called the GPT-5-Codex, takes “thinking” time more dynamically than previous models and can spend seconds to seven hours on any time on coding tasks. As a result, it performs better in proxy coding benchmarks.

Now, the new model is being launched in Codex products, allowing access to all Chatgpt Plus, Pro, Business, Edu and Enterprise users through Terminal, IDE, GITHUB or CHATGPT. Openai said it plans to use the model for API customers in the future.

The update is part of Openai’s efforts to make Codex more competitive with other AI-encoding products, such as Claude Code, Anysphere’s cursor, or Microsoft’s GitHub Copilot. Due to the high demand for users, the market for AI encoding tools has become increasingly crowded last year. cursor More than $500 million in ARR Earlier in 2025, similar code edits Windsurf were Chaotic acquisition attempts This splits its team between Google and cognition.

Openai says GPT-5-Codex is in SWE bench verifieda benchmark for measuring agent coding capabilities, and a benchmark for measuring the performance of code refactoring tasks from large, established repositories.

The company also said it trained GPT-5-Codex for code reviews and asked to experience comments from software engineers evaluating models. Engineers reportedly found that GPT-5-Codex submitted fewer false comments, while adding more “high impact comments”.

In the briefing, Alexander Embiricos, Openai’s code product head, said that most of the performance is attributed to the dynamic “thinking ability” of GPT-5-Codex. Users may be familiar with the GPT-5 router in Chatgpt, which leads queries to different models depending on the complexity of the task. Embiricos says GPT-5-Codex works similarly, but does not have a router under the hood and can adjust how long it takes to perform tasks in real time.

Embiricos says this is an advantage over a router, which decides how much computing power and time it is used on the issue from the beginning. Instead, GPT-5-Codex can decide what it takes to take an hour and five minutes. Embiricos said in some cases he has seen the model take more than seven hours.