OpenAI today released GPT-5-Codex, a version of its flagship model purpose-built for autonomous coding agents. The model is now the default in the Codex CLI, IDE extension, cloud service, and GitHub code review integration, and is also accessible through the ChatGPT app to subscribers.

Trained with reinforcement learning on real-world engineering tasks — including building full projects from scratch, adding features and tests, debugging, and large-scale refactoring — GPT-5-Codex can operate independently for more than seven hours. It uses what OpenAI calls “adaptive reasoning”: for the simplest 10% of requests, the model uses 93.7% fewer tokens than GPT-5; for the most complex 10%, it spends roughly twice as long on reasoning, editing, and testing. In refactoring benchmarks, GPT-5-Codex achieved 51.3% accuracy versus GPT-5’s 33.9%, including successfully modifying a pull request across 232 files and 3,500 lines of code in the Gitea repository.

Code review capabilities have been strengthened: the model navigates repositories, analyzes dependencies, and runs tests. OpenAI reports that in evaluations on recent open-source commits, GPT-5-Codex produced more accurate and higher-value review comments. The release puts OpenAI squarely against Anthropic’s Claude Code, which developers have leaned on for similar agentic workflows. With API access promised soon, the pitch is unambiguous: a coding agent you hand a task and walk away from.