GPT-5.2-Codex Release Notes

OpenAI has released GPT-5.2-Codex, a specialized version of GPT-5.2 optimized for agentic coding within the Codex environment. This model is designed to handle complex, real-world software engineering tasks, including large-scale refactors and migrations, while providing significantly enhanced cybersecurity capabilities.

Advanced Agentic Software Engineering

GPT-5.2-Codex improves upon the capabilities of GPT-5.2 and GPT-5.1-Codex-Max by focusing on long-horizon work and reliability in terminal environments. The model is designed to be a dependable partner for extended coding sessions where maintaining full context is critical.

Key Technical Improvements

  • Context Compaction: The model utilizes native compaction to remain token-efficient during reasoning while improving long-context understanding.
  • Tool Calling and Factuality: Improvements in reliable tool calling and factuality reduce errors during long-running tasks.
  • Environment Support: The model demonstrates increased effectiveness and reliability when operating in native Windows environments.
  • Vision Integration: Enhanced vision performance allows the model to accurately interpret technical diagrams, charts, UI surfaces, and screenshots, enabling the translation of design mocks into functional prototypes.

Performance Benchmarks

GPT-5.2-Codex achieves state-of-the-art performance on two primary benchmarks:

  • SWE-Bench Pro: Tests the model's ability to generate patches to solve realistic software engineering tasks within a code repository.
  • Terminal-Bench 2.0: Evaluates agentic performance in real terminal environments, including tasks such as setting up servers, training models, and compiling code.

Cybersecurity Capabilities and Dual-Use Risks

GPT-5.2-Codex represents a significant jump in cybersecurity capability compared to previous iterations. While it has not yet reached a "High" level of cyber capability according to OpenAI's Preparedness Framework, the lab is treating its deployment with caution due to the potential for dual-use risks.

Real-World Application in Vulnerability Research

OpenAI highlights a case where GPT-5.1-Codex-Max was used by a security researcher, Andrew MacPherson, to discover previously unknown vulnerabilities in React Server Components. By using the Codex CLI and iterative prompting—including setting up local test environments and using fuzzing to probe systems—the researcher was able to responsibly disclose new vulnerabilities to the React team on December 11, 2025.

Safety and Deployment Framework

To balance accessibility with safety, OpenAI is implementing the following measures:

  • Gradual Rollout: The model is available in all Codex surfaces for paid ChatGPT users, with API access planned for the coming weeks.
  • Additional Safeguards: New safeguards have been integrated into the model and product, as detailed in the system card.
  • Trusted Access Pilot: An invite-only program is being piloted for vetted security professionals and organizations. This program provides more permissive models and advanced capabilities to authorized defenders to accelerate defensive security work, such as malware analysis and infrastructure stress testing.

Summary of Availability

Access Tier Availability
Paid ChatGPT Users Available now in all Codex surfaces
API Users Coming in the coming weeks
Vetted Professionals Invite-only trusted access pilot for defensive cyber work

Sources