GPT-5.1-Codex-Max System Card

OpenAI has introduced GPT-5.1-Codex-Max, a frontier agentic coding model designed for complex software engineering, mathematics, and research tasks. The model is distinguished by its ability to operate across multiple context windows using a process called compaction, allowing it to maintain coherence over millions of tokens in a single task.

Model Architecture and Training

GPT-5.1-Codex-Max is built upon an update to OpenAI's foundational reasoning model. It was natively trained to handle agentic tasks across several domains, including software engineering, medicine, and computer use. To optimize its performance on real-world software engineering workflows, the model was trained on tasks such as pull request (PR) creation, code review, frontend coding, and technical Q&A.

Safety and Mitigation Strategies

OpenAI has implemented a comprehensive suite of safety measures to protect against misuse and vulnerabilities. These mitigations are divided into two primary categories:

Model-Level Mitigations

Model-level protections include specialized safety training specifically targeted at preventing the model from performing harmful tasks and resisting prompt injections.

Product-Level Mitigations

Product-level safeguards include agent sandboxing and configurable network access to ensure the model's agentic capabilities are executed in a controlled environment.

Preparedness Framework Evaluation

GPT-5.1-Codex-Max was evaluated using OpenAI's Preparedness Framework to assess its potential risks. The results are as follows:

  • Cybersecurity: The model is described as very capable in the cybersecurity domain, but it does not currently reach the "High" capability threshold. OpenAI expects that future models will likely cross this threshold as capabilities continue to increase.
  • Biology: The model is treated as having "High" capability in biology, and is therefore deployed with the same safeguards used for GPT-5.
  • AI Self-Improvement: The model does not reach the "High" capability threshold for AI self-improvement.

Sources