GPT 5.3 Codex is here and it debugged its own training says OpenAI
OpenAI released GPT-5.3-Codex, a new coding model with improved reasoning and professional knowledge capabilities that operates 25 percent faster than its predecessor. Anthropic launched rival model Claude Opus 4.6 on the same day. Earlier this week, OpenAI introduced the Codex app for macOS to manage multiple AI agents simultaneously.
OpenAI describes GPT-5.3-Codex as its first model instrumental in creating itself. The company explains that developers leveraged early versions of the model during its development process. These versions assisted in specific technical tasks essential to building the final product.
In a blog post announcing the model, OpenAI states: “The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations — our team was blown away by how much Codex was able to accelerate its own development.” Debugging training involves identifying and correcting errors in the model’s learning algorithms. Managing deployment covers coordinating the rollout of the model across servers and infrastructure. Diagnosing test results requires analyzing performance metrics from benchmarks and evaluations to refine outputs.
Anthropic recently made a comparable assertion about its Clade Cowork model. Engineers at both OpenAI and Anthropic report that artificial intelligence now performs almost all of their coding work. This shift indicates heavy reliance on AI tools for software development tasks at these organizations.
The OpenAI blog post further details the model’s expanded abilities. It specifies: “With GPT‑5.3‑Codex, Codex goes from an agent that can write and review code to an agent that can do nearly anything developers and professionals can do on a computer.” Writing code entails generating functional scripts and programs. Reviewing code includes checking for bugs, inefficiencies, and adherence to standards. The broader scope encompasses tasks like data analysis, file management, and interface interactions typically handled by human experts on computing devices.
The announcements highlight self-improving models as real-world examples relevant to discussions of the technological singularity. These discussions define the singularity as a tipping point where technology becomes self-improving, leading to an uncontrolled explosion of technological advancement.