技术规格

GPT-5.1 Codex

GPT-5.1 Codex is a specialized large language model from OpenAI, engineered for high-fidelity software development and agentic coding workflows. Built upon the GPT-5.1 foundation, this variant is optimized for long-horizon engineering tasks where maintaining state and coherence across complex repositories is essential. Unlike general-purpose models, Codex is specifically tuned to operate as an autonomous agent within development environments, capable of performing multi-file refactoring, autonomous debugging, and test-driven development cycles that may persist for extended periods.

The architecture utilizes a dense transformer configuration with multi-head attention (MHA), supporting an extensive context window of up to 400,000 tokens. A primary innovation in this series is the implementation of a session compaction mechanism. When the interaction nears the context limit, the model prunes its conversation history while preserving critical architectural details and logic, effectively allowing it to sustain coherence over tasks that would otherwise exceed standard hardware constraints. The model also features a dynamic reasoning engine, where developers can adjust the computational effort through API parameters to balance latency with the depth of technical analysis required for a specific problem.

Functionally, GPT-5.1 Codex integrates natively with modern development toolchains via the Responses API. It is equipped with specialized tools such as apply_patch for reliable code modification and a shell interface for executing terminal commands within a controlled environment. This makes the model particularly effective for complex software engineering pipelines, including dependency management, environment setup, and large-scale architectural migrations. Its training objective prioritizes precise adherence to developer instructions and the generation of clean, production-ready code, reducing common issues like sycophancy or hallucinated syntax in technical responses.