GPT-5.1-Codex is engineered as an agentic coding companion built for long-running, multi-step development workflows rather than simple code completion. It introduces a context compaction technique that lets it work coherently across multiple context windows, effectively handling million-token codebases in a single task. The model accepts both text and image inputs, enabling it to reason across code, UI states, architecture diagrams, and design comps within the same workflow. Its repo-aware intelligence understands full repositories, supporting cohesive refactors and test automation, while model-guided loops retain state and context across extended interactions for truly asynchronous execution of long-running coding tasks.
Post-training incorporates reinforcement learning and supervised fine-tuning to cultivate agentic behavior, with reasoning effort levels that let developers trade latency for code quality depending on task complexity. The xhigh reasoning level achieved 77.9% on SWE-bench Verified while using 30% fewer thinking tokens, and Terminal Bench 2.0 scores of 58.1% outpaced competing models from Gemini and Anthropic. An independent METR evaluation found the model posed low risk for AI R&D automation and rogue replication threats, with OpenAI observing the system operate autonomously for over 24 hours continuously—iterating through code and fixing failures without human intervention. These capabilities make it well-suited for teams deploying autonomous development agents at scale.
