GPT-5.1-Codex represents a shift from code completion tools toward autonomous software development agents. The model introduces context compaction—a technique that lets it operate coherently across multiple context windows and process millions of tokens within a single task, far beyond earlier token limits. It is designed for steerability, front-end development, and interactive workflows, accepting both text and image inputs to reason across code, UI states, and architecture diagrams. Repo-aware intelligence enables cohesive refactoring and test automation across entire codebases, while asynchronous execution allows long-running tasks to retain state and context across interactions without manual resets.

The model ships with multiple reasoning levels, including a new xhigh effort setting that achieved 77.9% on SWE-bench Verified while using 30% fewer thinking tokens than its predecessor. On Terminal Bench 2.0, it scores 58.1%, surpassing Gemini at 54.2% and Sonnet at 42.8%. OpenAI observed Codex-Max working continuously for over 24 hours, iteratively implementing features, running tests, and fixing failures without human intervention. Independent evaluation by METR found it represents a low-risk incremental improvement with no risk-critical changes to affordances or architecture, projecting continued on-trend advancement over a six-month horizon. This combination of token efficiency, autonomous endurance, and independent safety validation positions the model for sustained multi-day software engineering tasks.