Last month, OpenAI announced the most advanced artificial intelligence model GPT-5, which includes significant improvements in logical reasoning, coding, and many other areas. The company stated that this model has made great progress in “agent-like” capabilities. GPT-5 achieved a success rate of 74.9% in the SWE-bench Verified test and 88% in the Aider Polyglot test.
GPT-5-Codex: A New Model Specialized for Coding
Today, OpenAI introduced GPT-5-Codex, the new version optimized for coding. This model was developed specifically to be more effective in real-world software engineering tasks. It demonstrates higher performance in creating new projects from scratch, adding features and tests to existing projects, and large-scale code refactoring.
Higher Success in Tests
In SWE-bench Verified comparisons, GPT-5 High showed a success rate of 72.8%, while GPT-5-Codex reached 74.5%. In code editing tasks, it made a significant difference by surpassing GPT-5’s 33.9% score with a 51.3% success rate.
Ability to Work Independently for Long Periods
One of the most notable features of GPT-5-Codex is its ability to determine the thinking time according to the complexity of the task. Internal tests conducted by OpenAI observed that the model could work continuously for over 7 hours on complex tasks.
Visual-Assisted Coding and Mobile Compatibility
The new model is especially more successful in mobile websites and frontend development. It can accept images as input via the cloud, track its progress visually, and share screenshots of its work.
Smarter Code Review
GPT-5-Codex also stands out in code review processes. It can match the intent of a pull request with actual changes, perform logical deductions across the entire codebase, and verify the expected behavior by running code/tests. OpenAI announced that this model is now used in internal code reviews and detects hundreds of issues daily.
Made Available
The GPT-5-Codex model is accessible within the ChatGPT Plus, Pro, Business, Edu, and Enterprise packages for code reviews via Codex CLI, IDE extension, web, mobile, and GitHub. Developers using the Codex CLI with an API key have been informed that the model will soon be accessible via API as well.
