OpenAI has officially unveiled GPT-5-Codex, a significant upgrade to its AI coding agent, Codex. Announced on Monday, this new model is designed to revolutionize how AI handles coding tasks by employing a novel dynamic “thinking” mechanism. Unlike its predecessors, GPT-5-Codex can allocate its processing time much more flexibly, dedicating anywhere from a few seconds to an impressive seven hours to a single coding problem. This adaptive approach has led to notable improvements in its performance across various agentic coding benchmarks.

The rollout of GPT-5-Codex is currently underway for users of Codex products, which are accessible through various platforms including a terminal, IDEs, GitHub, and ChatGPT. Specifically, the update is being made available to all ChatGPT Plus, Pro, Business, Edu, and Enterprise subscribers. OpenAI has also indicated future plans to extend availability to its API customers, broadening the reach of this advanced coding assistant.

This initiative by OpenAI is a strategic move to bolster Codex's competitiveness within the rapidly expanding market for AI coding tools. The past year has seen an intense surge in user demand, leading to a crowded landscape filled with strong contenders such as Claude Code, Anysphere’s Cursor, and Microsoft’s GitHub Copilot. In this competitive environment, GPT-5-Codex demonstrates superior capabilities, outperforming the standard GPT-5 model on the SWE-bench Verified benchmark, which rigorously measures agentic coding abilities. Furthermore, it excels in benchmarks designed to assess performance on complex code refactoring tasks sourced from large, established code repositories.

Beyond its core coding proficiency, OpenAI has also optimized GPT-5-Codex for conducting thorough code reviews. The company conducted evaluations where experienced software engineers assessed the model's review comments. The feedback highlighted that GPT-5-Codex produced significantly fewer incorrect comments while simultaneously delivering a higher proportion of “high-impact comments,” thereby enhancing the quality and efficiency of the code review process.

During a briefing, Alexander Embiricos, OpenAI’s Codex product lead, attributed much of the model's enhanced performance to its innovative dynamic “thinking abilities.” Embiricos drew parallels to GPT-5’s router in ChatGPT, which intelligently directs queries to different models based on their complexity. However, he clarified that GPT-5-Codex operates without an explicit router under the hood, instead possessing the capability to adjust the duration it dedicates to a task in real-time. This is a key advantage, as it allows the model to re-evaluate and extend its effort on a problem even after several minutes, a stark contrast to fixed-time allocation systems. Embiricos revealed that in some instances, the model has been observed to spend upwards of seven hours on particularly challenging tasks.

The broader market context underscores the significance of this release. Companies like Cursor have achieved substantial financial milestones, surpassing $500 million in Annual Recurring Revenue (ARR) earlier in 2025. Another similar code editor, Windsurf, experienced a turbulent acquisition attempt, resulting in its team being divided between tech giants Google and Cognition. OpenAI's investment in GPT-5-Codex reflects the critical importance of innovation and performance in this dynamic and rapidly evolving sector.