OpenAI Launches GPT-5-Codex with 51% Accuracy on Code Refactoring, Operates Autonomously for 7+ Hours
Summary
OpenAI releases GPT-5-Codex, a specialized AI that achieves 51% accuracy on code refactoring tasks and operates autonomously for over seven hours, dramatically outperforming standard GPT-5 while using 94% fewer tokens on simple coding requests.
Key Points
- OpenAI releases GPT-5-Codex, a specialized version of GPT-5 optimized for complex software engineering tasks including large-scale code refactoring and extended code review workflows that can operate autonomously for over seven hours
- The model features adaptive reasoning that adjusts computation time based on task complexity, achieving 51.3% accuracy on refactoring tasks compared to 33.9% for GPT-5, while using 93.7% fewer tokens on simple requests
- GPT-5-Codex is now available through the Codex CLI and IDE extension as the default in Codex's cloud service, with early users reporting it successfully catches backward compatibility issues and produces high-quality tested code