Openai announced on Monday that it will release a new version of GPT-5 to AI coding agent Codex. The company says a new model called the GPT-5-Codex is spending more dynamically “thinking” time than its previous models, allowing it to be spent anywhere from seconds to seven hours on coding tasks. As a result, agent coding benchmarks provide better performance.
The new model is deployed in Codex products that can be accessed via terminal, IDE, GitHub, or ChatGpt. Openai says it plans to make the model available to API customers in the future.
This update is part of Openai’s efforts to make Codex more competitive with other AI coding products such as Claude Code, Anysphere’s cursor, and Microsoft’s Github Copilot. The market for AI coding tools has been much more crowded last year as a result of strong user demand. Cursors surpassed $500 million in early 2025, and Windsurf, a similar code editor, was the subject of a chaotic attempt at acquiring a team split between Google and Cognition.
According to Openai, GPT-5-Codex outperforms the SWE bench validation, benchmark measurement agent coding capabilities, and benchmark measurement performance for code refactoring tasks from large established repositories.
The company also said it had trained GPT-5-Codex to carry out code reviews and asked experience software engineers to rate review comments on the model. Engineers reportedly found that GPT-5-Codex adds “high impact comments” when it comes to less false comments.
During the briefing, Alexander Embiricos, lead of Openai’s Codex product, said the increase in performance was due to the dynamic “thinking ability” of the GPT-5-Codex. Users may be familiar with TASKGPT’s GPT-5 routers. This directs queries to different models based on the complexity of the task. Embiricos said the GPT-5-Codex works similarly, but does not have a router under the hood, allowing you to adjust the time to tackle the task in real time.
Embiricos says this is an advantage over routers. This is to determine the computing power and time to use in the first problem. Instead, the GPT-5-Codex can be determined after 5 minutes of the problem that requires an additional hour to be spent. Embiricos said he saw the model take more than seven hours in some cases.
TechCrunch Events
San Francisco
|
October 27th-29th, 2025
