OpenAI releases GPT-5.3-Codex, the first AI model that helped build itself

OpenAI has unveiled GPT-5.3-Codex, positioning it as their most advanced agentic coding model. The system merges enhanced coding abilities from GPT-5.2-Codex with the analytical and professional competencies of GPT-5.2, delivering performance that's 25% quicker than previous iterations. This allows the model to tackle extended tasks that require investigation, tool integration, and sophisticated implementation.

The standout feature of this launch is GPT-5.3-Codex's role in its own development. OpenAI's teams deployed preliminary versions to troubleshoot training processes, handle infrastructure deployment, and analyze evaluation outcomes. Research staff utilized it to observe training progressions, assess interaction quality, and suggest improvements. Engineering teams relied on it to refine system performance, detect rendering issues, and diagnose cache problems. Throughout the launch phase, the model managed GPU resource allocation to handle traffic fluctuations and preserve response times.

GPT-5.3-Codex demonstrates leading results on multiple industry evaluations including SWE-Bench Pro and Terminal-Bench, while also performing strongly on OSWorld and GDPval. In SWE-Bench Pro, which evaluates practical software engineering across four programming languages, the model achieves top-tier scores with reduced token consumption compared to predecessors. It substantially surpasses earlier benchmarks on Terminal-Bench 2.0, which assesses command-line proficiency essential for coding agents. On OSWorld, which tests agents on productivity assignments within visual desktop environments, GPT-5.3-Codex shows considerably stronger computer operation capabilities than prior GPT versions.

A new interactive element allows users to guide the model during active work sessions. Rather than submitting requests and awaiting completion, users can pose questions, explore different methodologies, and redirect the model's approach in real time while preserving full context. This steering functionality is accessible through the Codex app settings.

OpenAI tested the model's sustained performance by having it autonomously develop two sophisticated games across millions of tokens using basic instructions such as "fix the bug" or "improve the game." The model produced a racing title and an underwater exploration game featuring reef discovery, marine life collection, and oxygen and pressure mechanics with environmental challenges. Both applications are operational and illustrate the model's capacity for projects spanning multiple days.

The deployment incorporates extensive protective measures. GPT-5.3-Codex represents OpenAI's first model designated as high capability for cybersecurity functions under their Preparedness Framework, and their first trained explicitly to detect software vulnerabilities. According to the announcement: "While we don't have definitive evidence it can automate cyber attacks end-to-end, we're taking a precautionary approach and deploying our most comprehensive cybersecurity safety stack to date." These protections encompass safety instruction, continuous automated oversight, restricted access protocols for advanced features, and enforcement systems incorporating threat intelligence. OpenAI is initiating Trusted Access for Cyber, an experimental initiative to support defensive security research, alongside allocating $10 million in API credits toward cyber defense efforts, particularly for open source projects and essential infrastructure systems.

The model is currently accessible to paid ChatGPT subscribers across all Codex platforms including the application, command-line interface, IDE extensions, and web interface. API availability is being prepared with appropriate security considerations. Additionally, the model performs competitively on GDPval, which evaluates capabilities across clearly defined professional assignments spanning 44 occupations, encompassing tasks like presentation development, spreadsheet creation, and various business deliverables.

Try Closer AI website builder to create AI-powered online experiences while these frontier models continue to evolve.