OpenAI upgrades Codex with a new version of GPT-5

OpenAI introduced Monday that it’s releasing a brand new model of GPT-5 to its AI coding agent, Codex. The corporate says its new mannequin, referred to as GPT-5-Codex, spends its “considering” time extra dynamically than earlier fashions and will spend wherever from a number of seconds to seven hours on a coding activity. In consequence, it performs higher on agentic coding benchmarks.

The brand new mannequin is now rolling out in Codex merchandise — which could be accessed by way of a terminal, IDE, GitHub, or ChatGPT — to all ChatGPT Plus, Professional, Enterprise, Edu, and Enterprise customers. OpenAI says it plans to make the mannequin accessible to API prospects sooner or later.

The replace is a part of OpenAI’s effort to make Codex extra aggressive with different AI coding merchandise, comparable to Claude Code, Anysphere’s Cursor, or Microsoft’s GitHub Copilot. The marketplace for AI coding instruments has develop into way more crowded within the final 12 months on account of intense person demand. Cursor surpassed $500 million in ARR earlier in 2025, and Windsurf, an identical code editor, was the topic of a chaotic acquisition try that noticed its crew break up between Google and Cognition.

OpenAI says that GPT-5-Codex outperforms GPT-5 on SWE-bench Verified, a benchmark measuring agentic coding skills, in addition to a benchmark measuring efficiency on code refactoring duties from massive, established repositories.

The corporate additionally says it educated GPT-5-Codex for conducting code evaluations and requested expertise software program engineers to guage the mannequin’s assessment feedback. The engineers reportedly discovered GPT-5-Codex to submit fewer incorrect feedback, whereas including extra “high-impact feedback.”

In a briefing, OpenAI’s Codex product lead Alexander Embiricos stated that a lot of the elevated efficiency was because of GPT-5-Codex’s dynamic “considering skills.” Customers could also be conversant in GPT-5’s router in ChatGPT, which directs queries to totally different fashions primarily based on the complexity of a activity. Embiricos stated GPT-5-Codex works equally however has no router underneath the hood and may alter for the way lengthy to work on a activity in actual time.

Embiricos says this is a bonus in comparison with a router, which decides how a lot computational energy and time to make use of on an issue on the outset. As an alternative, GPT-5-Codex can resolve 5 minutes into an issue that it must spend one other hour. Embiricos stated he’s seen the mannequin take upward of seven hours in some instances.