OpenAI on Thursday launched GPT-5.3-Codex, a new artificial intelligence model built for advanced agentic coding tasks. According to the San Francisco–based company, the model takes Codex’s capabilities to a new level and can create complex applications and video games from scratch.
The company described GPT-5.3-Codex as its most capable agentic coding system so far. It can manage the entire software development process, debug large codebases, research technical requirements, and deploy changes end to end. Notably, it is also the first OpenAI model to play an active role in its own development.
Availability and Key Features
In a blog post, OpenAI said GPT-5.3-Codex is now available to users on all paid ChatGPT plans worldwide. Users can access it through mobile and desktop apps, the command-line interface (CLI), IDE extensions, and the web. The company plans to introduce API support soon.
The new model combines the coding strength of GPT-5.2-Codex with the reasoning ability and professional knowledge depth of GPT-5.2 into a single system. OpenAI said GPT-5.3-Codex is about 25% faster, making it more effective for research-driven and complex multi-step projects.
Unlike earlier versions, users can now guide the model while it is working. They can request progress updates, ask questions, suggest improvements, or discuss alternative approaches without the system losing context.
Faster Development With Self-Help
According to OpenAI, early versions of GPT-5.3-Codex supported the Codex team during development. The model helped debug training processes, manage deployments, and analyze evaluation results. The company said this self-assistance significantly accelerated development.
Performance Benchmarks
OpenAI also shared internal benchmark results. On SWE-Bench Pro, a challenging real-world software engineering test, GPT-5.3-Codex achieved 56.8% accuracy, outperforming GPT-5.2-Codex and GPT-5.2. On Terminal-Bench 2.0, its score rose to 77.3%, compared with 64.0% for the previous Codex version. In OSWorld-Verified, the model scored 64.7%, while GPT-5.2-Codex reached 38.2%.
Beyond Coding and Focus on Safety
GPT-5.3-Codex can generate complex web games from minimal prompts and iterate autonomously over millions of tokens. One demo showcased a fully functional racing game with tracks, items, and racers. The model also produces production-ready websites, automatically adding features such as discount sections and testimonial carousels.
Beyond coding, the model supports the full product lifecycle, including writing PRDs, editing content, conducting user research, creating slide decks, analyzing spreadsheets, and monitoring systems.
OpenAI said it has placed strong emphasis on safety. GPT-5.3-Codex is the first model classified as “High Capability” under the company’s Preparedness Framework for cybersecurity tasks. It includes advanced safeguards such as specialized safety training, automated monitoring, controlled access, and threat intelligence enforcement.