GPT-5.1-Codex-Max System Card

OpenAI released the System Card for GPT-5.1-Codex-Max, its new frontier agentic coding model, which OpenAI says is its first model natively trained with compaction, able to coherently process millions of tokens across multiple context windows in a single task. The model was evaluated under OpenAI's Preparedness Framework. It is highly capable in cybersecurity but does not reach High capability, and OpenAI expects the trend of rapid capability improvement to continue, with a model crossing that threshold in the near term. In the biological domain it is treated as High capability, using the same safeguards as GPT-5. AI self-improvement did not reach High. Mitigations include model-level measures such as dedicated safety training against harmful tasks and prompt injection, as well as product-level measures such as the Agent sandbox and configurable network access.