GPT-5.3-Codex Released

GPT-5.3-Codex is the most powerful agentic coding model, combining reasoning and expert knowledge capabilities beyond just coding performance. It boasts 25% faster speed compared to the existing GPT-5.2-Codex, and can carry out complex, long-running tasks that involve research and tool use. Users can collaborate with the model by giving real-time instructions while it is working, without it losing context.

In particular, this model is notable for having been directly involved in its own development process. The Codex team used early versions of the model to perform training debugging, deployment management, and test result diagnosis, dramatically speeding up model development.

Its key capabilities are as follows:

• Coding and terminal control: Achieved industry-leading performance on SWE-Bench Pro and Terminal-Bench 2.0. It performs more sophisticated tasks while using fewer tokens than previous models.
• Web development: Can autonomously build complex games or apps over the course of several days. It understands user intent more deeply, producing polished landing pages from simple prompts alone.
• Expanding into professional work: Beyond writing code, it supports the entire software lifecycle and expert knowledge work, including debugging, deployment, monitoring, PRD writing, and data analysis.