openai upgrades codex with a new version of gpt-5

OpenAI Announces GPT-5-Codex for Enhanced Coding

OpenAI revealed on Monday the release of a new iteration of GPT-5, specifically tailored for its AI coding assistant, Codex.

The company highlights that this updated model, designated GPT-5-Codex, allocates its processing time more flexibly compared to its predecessors.

It can dedicate anywhere from a few seconds to as much as seven hours to a single coding challenge, resulting in improved performance on agentic coding evaluations.

Availability and Access

The rollout of this new model is currently underway across all Codex products.

Access is available through various interfaces, including a terminal, integrated development environments (IDEs), GitHub, and ChatGPT.

Currently, GPT-5-Codex is accessible to subscribers of ChatGPT Plus, Pro, Business, Edu, and Enterprise plans.

OpenAI intends to extend access to its API customers in the near future.

Competitive Landscape

This update represents OpenAI’s ongoing efforts to strengthen Codex’s position within the increasingly competitive AI coding tool market.

Notable competitors include Claude Code, Anysphere’s Cursor, and Microsoft’s GitHub Copilot.

Demand for AI-powered coding solutions has surged in the past year, leading to a more crowded marketplace.

For example, Cursor exceeded $500 million in annual recurring revenue (ARR) earlier in 2025.

Furthermore, Windsurf, a comparable code editor, was the focus of a complex acquisition process that ultimately divided its team between Google and Cognition.

Performance Improvements

OpenAI reports that GPT-5-Codex demonstrates superior performance compared to GPT-5 on SWE-bench Verified.

This benchmark specifically assesses agentic coding capabilities.

The new model also excels in benchmarks evaluating code refactoring performance within large, established code repositories.

Enhanced Code Review Capabilities

The company also focused on training GPT-5-Codex for conducting thorough code reviews.

Experienced software engineers were tasked with evaluating the model’s feedback on code.

Their assessment indicated that GPT-5-Codex generated fewer inaccurate comments while simultaneously providing a greater number of “high-impact” insights.

Dynamic “Thinking” Abilities

During a briefing, Alexander Embiricos, OpenAI’s Codex product lead, explained that the performance gains are largely attributable to GPT-5-Codex’s dynamic processing approach.

Users familiar with GPT-5’s routing system in ChatGPT, which directs inquiries to different models based on task complexity, will recognize a similar principle.

However, GPT-5-Codex operates without a traditional router, allowing it to adjust the time allocated to a task in real-time.

Advantages Over Traditional Routing

Embiricos emphasized that this approach offers an advantage over conventional routing systems.

Routers typically determine the necessary computational resources and time commitment at the beginning of a problem.

In contrast, GPT-5-Codex can reassess its needs mid-process, potentially extending its work duration by an additional hour after just five minutes of analysis.

Embiricos noted instances where the model has spent over seven hours on particularly complex tasks.