OpenAI Unveils GPT-5.2-Codex, Its Most Advanced AI Programming Agent

OpenAI has released GPT-5.2-Codex, an AI agent programming model designed for complex software engineering tasks. The company describes it as its most powerful AI agent to date, featuring enhanced context compression, improved long-cycle task processing, and stronger performance in large code changes.

GPT-5.2-Codex is an optimized version of GPT-5.2, offering significant advancements in several areas. These include better handling of extensive code modifications like refactoring and migration, enhanced programming capabilities within native Windows environments, and its most robust cybersecurity features to date. OpenAI CEO Sam Altman reportedly stated that OpenAI has adopted the model internally with positive results.

Sam Altman, CEO of OpenAI, in a professional setting.

In benchmark tests, GPT-5.2-Codex surpassed previous models, including 5.1-Codex-Max, GPT-5.2, and GPT-5.1, across software engineering and terminal testing scenarios. OpenAI's blog emphasizes the model's high level of cybersecurity capabilities.

Enhanced Programming and Long-Running Task Support

GPT-5.2-Codex combines the "professional task processing capabilities" of GPT-5.2 with the AI agent programming and terminal operation strengths of 5.1-Codex-Max. This integration has led to substantial improvements in long-context understanding, tool invocation, factual accuracy, and native context compression. The model can support stable, long-running programming tasks and optimize token usage during inference.

The model achieved new state-of-the-art (SOTA) records on industry-standard benchmarks such as SWE-Bench Pro and Terminal-Bench 2.0, showing approximately a 6% performance improvement over 5.1-Codex. These benchmarks assess an agent's ability to manage diverse tasks in real terminal environments. Its agent programming performance in native Windows environments has also seen significant improvement, building on the capabilities introduced by GPT-5.1-Codex-Max.

These enhancements enable GPT-5.2-Codex to work on large codebases for extended periods while maintaining full context. This allows it to reliably complete complex tasks such as large-scale refactoring, code migration, and feature development, even when plans change or attempts fail midway.

Holographic code and architectural diagrams projected above a developer's laptop.

Furthermore, GPT-5.2-Codex demonstrates improved "vision," allowing it to more accurately interpret screenshots, technical diagrams, charts, and various user interface (UI) elements. It can also read design drafts and convert them into functional prototypes, which developers can then refine collaboratively.

Cybersecurity Capabilities and Real-World Application

OpenAI identifies a clear progression in its AI models' cybersecurity capabilities. GPT-5-Codex marked the first significant improvement, followed by GPT-5.1-Codex-Max, and now GPT-5.2-Codex represents the third leap. The company anticipates continued evolution, with each generation potentially reaching the "high" level of cybersecurity defined in its "Preparedness Framework," although GPT-5.2-Codex has not yet attained this level.

The practical application of OpenAI's agent programming model in cybersecurity was demonstrated by Andrew MacPherson, Chief Security Engineer at Privy. Following the React team's disclosure of three security vulnerabilities in React Server Components on December 11, MacPherson used GPT-5.1-Codex-Max + Codex CLI and other programming agents to investigate.

After initial attempts with zero-shot learning and iterative prompting proved unsuccessful, he guided Codex to follow standard defensive security procedures: setting up a local test environment, analyzing potential attack surfaces, and injecting anomalous inputs through fuzzing. While attempting to reproduce an existing React vulnerability, Codex identified anomalous behavior that led to the discovery of a previously unknown critical React vulnerability within a week. This vulnerability was subsequently disclosed to the React team. This case highlights the potential of advanced AI systems to accelerate defensive security research.

Digital shield of code protecting server racks in a data center, symbolizing cybersecurity.

Availability and Developer Feedback

As of today, all paying users can access GPT-5.2-Codex, with the API expected to become available in the coming weeks.

Developer feedback on GPT-5.2-Codex has been mixed. One developer reported that the model failed to simulate vehicle and traffic light interactions. However, others noted its ability to generate intricate animation effects, similar to Gemini 3 Flash and Pro, and its strong performance in creating a Counter-Strike game.

OpenAI views the release of GPT-5.2-Codex as a significant step for AI in real-world software development and cybersecurity. It aims to assist developers with complex and time-consuming tasks and provide enhanced tool support for cybersecurity research.