OpenAI introduces GPT-5.2 Codex for coding

OpenAI has introduced GPT-5.2-Codex, a model specialized in programming and designed to work for longer on real-world, complex projects. The company describes it as its most advanced model for agentic software engineering tasks, meaning it can plan and execute multiple steps instead of simply answering questions.

The model is based on GPT-5.2 and optimized for Codex. One of its improvements is context compaction, a technique that summarizes and reorganizes relevant information when a task becomes lengthy. This helps it maintain context during work such as a system migration or a full refactoring, which means reorganizing code without changing what it does.

It also performs better in Windows environments and on tasks that affect an entire project, not just one isolated file. In practice, that points to work such as updating an old application, moving it to another technology, or modifying many parts of the code while keeping its dependencies under control.

More capability, with security limits

The announcement includes a safety document explaining how OpenAI evaluated and protected the model. The measures are divided into two levels:

• Specialized training to reduce its ability to assist with harmful tasks and resist prompt injection attacks, which try to make the model ignore its rules.
• Product controls, such as running the agent in an isolated environment and allowing its access to the internet and other networks to be configured.

These controls matter because a model that can edit files, execute actions, and work for hours has more room to cause harm if it receives malicious instructions or accesses systems without restrictions. Isolation limits what it can touch, while configurable network access reduces the connections it can make.

What OpenAI says about cybersecurity

GPT-5.2-Codex shows significantly stronger cybersecurity capabilities, according to OpenAI's evaluation. Even so, it does not reach the High level of cybersecurity capability within the Preparedness Framework, the system the company uses to measure risks in sensitive areas.

OpenAI warns that models' capabilities in this field are increasing rapidly and expects them to cross that threshold in the near future. It does not present that milestone as already achieved with this model.

In biology, the model is classified as having High capability, like other recent models. It is therefore deployed with the corresponding safeguards from the GPT-5 family. By contrast, it does not reach the High level for AI system self-improvement.

For you, the significance is that Codex is becoming more like a technical collaborator capable of maintaining long-running tasks and modifying entire projects, rather than simply suggesting code snippets. That could speed up migrations and maintenance, but it also makes it more important to review its changes, limit its permissions, and control which networks and files it can use. The next point to watch will be whether these improvements translate into reliable results during extended work, especially as its cybersecurity capabilities continue to grow.