According to AIbase, a milestone has been reached in the field of intelligent code generation today: GPT-5.2-Codex has been officially released. As the most advanced agent coding model currently available, GPT-5.2-Codex is deeply specialized and optimized based on the GPT-5.2 architecture, integrating the terminal operation expertise of GPT-5.1-Codex-Max, with the aim of overcoming complex software engineering and cybersecurity challenges.
The model has achieved a significant leap in long-term task execution. By introducing native context compaction technology, GPT-5.2-Codex can more efficiently and accurately handle large-scale code refactoring and system migration. In multiple authoritative benchmark tests, its performance has set industry records: it achieved an accuracy rate of 56.4% in SWE-Bench Pro, which evaluates real-world software engineering tasks, and an impressive 64.0% accuracy rate in Terminal-Bench 2.0, which simulates real terminal environments. Additionally, the model has seen targeted enhancements for Windows environments and can accurately interpret technical diagrams and user interface screenshots during the coding process.
In the highly anticipated field of cybersecurity, GPT-5.2-Codex has demonstrated exceptional potential as a defensive assistant. AIbase noted that its predecessor, GPT-5.1-Codex-Max, had previously helped security researcher Andrew MacPherson identify and disclose three important security vulnerabilities in the React framework within a week. Building on this, GPT-5.2-Codex has further enhanced its logical reasoning and fuzz testing capabilities. Although, according to the official "readiness framework" assessment, it has not yet reached a "high" risk level, it has shown professional-level capabilities in capture-the-flag (CTF) competitions.
To balance technological accessibility and potential dual-use risks, developers have implemented a multi-layered deployment strategy. Starting today, GPT-5.2-Codex will be available first in Codex CLI, IDE extensions, and cloud environments for paying ChatGPT users. For highly sensitive network defense needs, the official has also launched a "Trusted Access" pilot program, a controlled mechanism limited to invitations, aimed at providing targeted support to approved professional security organizations. Through this gradual release of capabilities, GPT-5.2-Codex will deeply empower global developers throughout the entire engineering lifecycle—from code repository navigation to automated vulnerability defense—while ensuring safety and compliance.
