GPT-5 Codex (high) is a specialized variant of OpenAI's GPT-5 foundation model, specifically optimized for high-complexity software engineering and agentic coding tasks. Released in September 2025, the model is designed to handle large-scale code refactoring, system-wide dependency analysis, and autonomous code reviews. It belongs to the "high" intelligence tier of the GPT-5 ecosystem, prioritizing deep reasoning and thoroughness over response speed.
One of the model's defining features is its adaptive reasoning capability, which allows it to dynamically adjust its internal computation based on the complexity of the programming task. This allows the model to perform "long-horizon" tasks, with some internal benchmarks demonstrating the ability to operate for several hours on single, complex pull requests. It achieved a 51.3% success rate on OpenAI's internal code refactoring evaluations, a significant improvement over the base GPT-5 model.
The model features a 400,000-token context window, enabling it to ingest and reason over extensive codebases or hundreds of source files simultaneously. This expanded context is paired with improved steerability and instruction following, intended to reduce the frequency of incorrect deletions or hallucinations common in earlier coding assistants during multi-file operations. GPT-5 Codex (high) is primarily integrated into developer tools such as the Codex CLI and specialized IDE extensions for agentic workflows.