2 hours ago
1
ARTICLE AD BOX
OpenAI has released GPT-5 Codex, a new variant of GPT-5 built for use as an automated coding agent. The model is designed to handle real development tasks like refactoring, generating tests, fixing bugs, and reviewing code. It's part of the updated Codex system, which runs inside terminals, IDEs, web environments, and GitHub.
According to OpenAI, GPT-5 Codex can take part in short, interactive coding sessions as well as long-running, independent tasks. In testing, it worked continuously for more than seven hours, iterating on complex problems and running multiple test cycles. The model adjusts its compute time depending on task difficulty - using fewer tokens than GPT-5 on simple tasks and more on harder ones.
Stronger results than GPT-5 on benchmarks
On standard benchmarks, GPT-5 Codex outperforms GPT-5. In the SWE-bench Verified dataset, it hits 74.5 percent accuracy compared to GPT-5's 72.8 percent. For refactoring tasks, GPT-5 Codex scores 51.3 percent, while GPT-5 manages just 33.9 percent. These benchmarks include challenges from established open-source projects written in languages like Python, Go, and OCaml.
The model was also trained specifically for code reviews. It checks commits, verifies dependencies, runs tests, and compares intended changes against actual diffs. According to OpenAI, experienced developers rated the output higher than GPT-5, with GPT-5 Codex providing fewer irrelevant or incorrect comments than other models.
Ad
THE DECODER Newsletter
The most important AI news straight to your inbox.
✓ Weekly
✓ Free
✓ Cancel at any time
Integration into development workflows
Codex is accessible through the command line, IDE extensions (including VS Code), and a cloud-based environment. In the CLI, developers can attach screenshots or diagrams to explain design choices. The system can generate to-do lists and connect to external tools such as web search or MCP servers.
The cloud version adds extra automation. It detects common setup scripts, installs missing dependencies, and can even launch its own browser to check visual output. Screenshots of results can be attached directly to tasks or pull requests.
GPT-5 Codex is included with ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. Higher-tier plans allow multiple projects per week, and Business and Enterprise users can purchase additional usage beyond their standard allocation. An API integration is planned but not available yet.
