OpenAI has launched GPT-5.3-Codex, a coding model that the company describes as its most capable agentic coding tool to date. The model runs 25% faster than its predecessor and combines advanced coding performance with reasoning capabilities in a single system.
According to OpenAI, GPT-5.3-Codex marks a significant milestone as the first model that helped create itself. The Codex team used early versions to debug training processes, manage deployment, and analyze test results. The model can now handle long-running tasks spanning multiple days while allowing users to interact and provide guidance without losing context.
The model sets new performance records on SWE-Bench Pro and Terminal-Bench, two industry benchmarks for coding and agentic capabilities. OpenAI reports that GPT-5.3-Codex achieves these results using fewer tokens than previous models, making operations more efficient.
Beyond writing code, the model supports various professional tasks including debugging, deployment monitoring, creating presentations, and analyzing spreadsheets. OpenAI demonstrated these capabilities by having GPT-5.3-Codex build two web-based games autonomously over millions of tokens. The model also shows improved understanding of user intent, generating more functional websites from simple prompts.
OpenAI classifies GPT-5.3-Codex as the first model with high capability for cybersecurity tasks under its Preparedness Framework. The company has implemented safety measures including automated monitoring and restricted access for advanced features. OpenAI is launching Trusted Access for Cyber, a pilot program for defense research, and committing $10 million in API credits for cybersecurity research.
