Addendum to GPT-5.2 System Card: GPT-5.2-Codex
OpenAI officially released GPT-5.2-Codex, a code-specialized variant of the GPT-5.2 series, achieving 89.3% pass@1 on HumanEval and 84.7% pass@1 on MBPP across Python, TypeScript, Go, and Rust.
Official Positioning and Technical Goals
GPT-5.2-Codex is a vertical enhancement of the GPT-5.2 architecture, optimized specifically for code generation and comprehension—emphasizing instruction following, cross-language coding, debugging assistance, and documentation synthesis. Its training data comprises public GitHub repositories (Star ≥ 500, up to Q3 2024), high-quality Stack Overflow Q&A, and official API documentation; no private codebases were included.
Key Performance Metrics
- Achieves 89.3% pass@1 on HumanEval (Python), +4.1 percentage points over GPT-5.2-base;
- Reaches 84.7% pass@1 on MBPP (multi-language programming problems), covering Python, TypeScript, Go, and Rust;
- Supports 128K-token context window and natively integrates with the Code Interpreter sandbox;
- Average inference latency is 320ms on an A100×8 cluster (2K input tokens, 512 output tokens).
Deployment and Availability
GPT-5.2-Codex is available exclusively via the OpenAI API (model ID: gpt-5.2-codex-202409) for commercial use; weights are not open-sourced, and no local deployment package has been released. Enterprise customers may apply for dedicated fine-tuning quotas, but fine-tuning is restricted to code-specific instruction tuning—no access to architectural modifications is granted.