OpenAI documents GPT-5.1-Codex-Max, an AI model

OpenAI has documented GPT-5.1-Codex-Max, an AI model specialized in programming that can work on very long tasks without losing track. The company presents it as its new model for software agents, systems that can plan and execute steps to complete work with fewer constant instructions.

The model is based on an update to its reasoning model and was trained on software engineering, mathematics, research, medicine, and computer-use tasks. In programming, its training examples include creating change requests, reviewing code, developing interfaces, and answering technical questions.

A task can span millions of tokens

GPT-5.1-Codex-Max is OpenAI's first model trained natively to work across multiple context windows. Context is the information a model can keep available while solving a task.

To do this, it uses a process called compaction, which summarizes and reorganizes previous information before continuing. According to OpenAI, this allows it to work coherently across millions of tokens within a single task, which is useful for large software projects or work that requires many stages.

In practice, it could review a repository, modify several files, check the result, and continue after a pause without having to start from scratch. This capability does not mean the model is autonomous without limits: its performance also depends on the tools, permissions, and environment in which it runs.

Safety in the model and the product

The technical report details safety measures applied at two levels:

• Model: specialized training to reduce dangerous responses and resist prompt injection attacks. These attacks try to make the system follow hidden instructions inside a file, a web page, or the code it is analyzing.
• Product: isolating the agent in a controlled environment known as a sandbox, with internet access that can be configured for each case.

This separation matters because a model capable of modifying code can also interact with files, services, and networks. Limiting those connections reduces the impact of an error or a malicious instruction.

What the risk assessment says

OpenAI evaluated GPT-5.1-Codex-Max using its preparedness framework for advanced models. The model shows high capability in cybersecurity, but does not reach the level defined as “High” in that category.

The company expects cybersecurity capabilities to keep increasing and warns that models could cross that threshold in the near future. In biology, GPT-5.1-Codex-Max is treated as having “High” capability, like other recent models, so it is deployed with the safeguards used for GPT-5.

By contrast, it does not reach the “High” level in autonomous AI systems improvement. This assessment measures whether a model can make a significant contribution to improving other artificial intelligence systems.

GPT-5.1-Codex-Max points in a clear direction: agents that do more than generate code snippets and can sustain long, structured jobs. What to watch now is whether that continuity leads to more reliable results and whether access controls can keep what the agent does in the real world under control.