OpenAI upgrades Codex with a new version of GPT-5

OpenAI announced Monday that it’s releasing a caller mentation of GPT-5 to its AI coding agent, Codex. The institution says its caller model, called GPT-5-Codex, spends its “thinking” clip much dynamically than erstwhile models, and could walk anyplace from a fewer seconds to 7 hours connected a coding task. As a result, it performs amended connected agentic coding benchmarks.

The caller exemplary is present rolling retired successful Codex products — which tin beryllium accessed via a terminal, IDE, GitHub, oregon ChatGPT — to each ChatGPT Plus, Pro, Business, Edu, and Enterprise users. OpenAI says it plans to marque the exemplary disposable to API customers successful the future.

The update is portion of OpenAI’s effort to marque Codex much competitory with different AI coding products, specified arsenic Claude Code, Anysphere’s Cursor, oregon Microsoft’s GitHub Copilot. The marketplace for AI coding tools has go overmuch much crowded successful the past year, arsenic a effect of aggravated idiosyncratic demand. Cursor surpassed $500 cardinal successful ARR earlier successful 2025 and Windsurf, a akin codification editor, was the taxable of a chaotic acquisition attempt that saw its squad divided betwixt Google and Cognition.

OpenAI says that GPT-5-Codex outperforms GPT-5 connected SWE-bench Verified, a benchmark measuring agentic coding abilities, arsenic good arsenic a benchmark measuring show connected codification refactoring tasks from large, established repositories.

(Credit: OpenAI)

The institution besides says it trained GPT-5-Codex for conducting codification reviews, and asked acquisition bundle engineers to measure the model’s reappraisal comments. The engineers reportedly recovered GPT-5-Codex to taxable less incorrect comments, portion adding much “high-impact comments.”

In a briefing, OpenAI’s Codex merchandise pb Alexander Embiricos said that overmuch of the accrued show was acknowledgment to GPT-5-Codex’s dynamic “thinking abilities.” Users whitethorn beryllium acquainted with GPT-5’s router successful ChatGPT, which directs queries to antithetic models based connected the complexity of a task. Embiricos said GPT-5-Codex works similarly, but has nary router nether the hood, and tin set for however agelong to enactment connected a task successful real-time.

Embiricos says this is an vantage compared to a router, which decides however overmuch computational powerfulness and clip to usage connected a occupation astatine the outset. Instead, GPT-5-Codex tin determine 5 minutes into a occupation that it needs to walk different hour. Embiricos said he’s seen the exemplary instrumentality upwards of 7 hours successful immoderate cases.