OpenAI launches GPT-5 Codex for real-world coding tasks

OpenAI has introduced GPT-5 Codex, a version of GPT-5 designed specifically for software development. The model has been trained on complex engineering tasks such as building applications, adding features, debugging, large-scale refactoring, and conducting code reviews.

Designed for Coding Workflows

GPT-5 Codex is now the default model for cloud tasks and code reviews in Codex, with developers able to choose it for local use through the Codex CLI and IDE extension. The model adapts to the complexity of the task, providing faster responses for smaller requests and extended reasoning for larger assignments. In testing, it has independently worked on tasks for more than seven hours, fixing test failures and delivering working implementations.

Benchmarks and Performance

OpenAI evaluated GPT-5 Codex across established benchmarks and real repositories:

• SWE-bench Verified (500 tasks): Accuracy at 72.8%, close to GPT-5’s 74.5%.

• Code Refactoring Tasks: Accuracy improved to 51.3%, compared to GPT-5’s 33.9%.

• Code Review Evaluations: Incorrect comments fell to 4.4% (down from GPT-5’s 13.7%), while high-impact review comments rose to 52.4% (from 39.4%).

The model also adjusts its token usage. On smaller tasks, it uses up to 93% fewer tokens than GPT-5, while on larger ones, it spends more time reasoning, editing, and testing code.

Updates to Codex Tools

Codex has been rebuilt to improve developer collaboration across environments. The CLI now supports attaching images such as screenshots or wireframes for context, tracking progress with to-do lists, and connecting to external systems with tool calls. Terminal UI improvements include clearer diffs, simplified approval modes, and better session management.

Codex is accessible in the terminal, IDEs, the web, GitHub, and the ChatGPT mobile app. It is available across ChatGPT Plus, Pro, Business, Edu, and Enterprise plans.

Safety Measures

The GPT-5 Codex system card outlines safeguards against unsafe use. Measures include specialised training to handle harmful tasks, protections against prompt injection, agent sandboxing, and configurable network access.

With these updates, Codex extends from a coding assistant into a more capable partner that can support real-world software engineering at scale.