OpenAI released GPT‑5.3‑Codex—a coding AI model that contributed to its own creation

OpenAI released GPT‑5.3‑Codex – a faster and more versatile code generation model

The company announced the release of its new code model, GPT‑5.3‑Codex.

Unlike its predecessor (GPT‑5.2‑Codex), it maintains the same effectiveness in writing programs while running 25 % faster and combines logical reasoning capabilities with the professional knowledge of the previous version.

Self‑development of the model – a key moment
OpenAI emphasized that GPT‑5.3‑Codex became the first model from the company that “helped itself” during its creation.

Developers used early iterations of the algorithm for:

* debugging training,
* managing its own deployment,
* diagnosing test results.

The team is amazed at how quickly and effectively the model accelerated its own growth.

From a code‑generator/checker to a universal assistant
The model has evolved from a simple code generator/validator into an agent capable of solving almost any task for developers and PC users.

It can:

* perform complex research using various tools,
* interact with the user in real time, allowing adjustments to the work direction.

New records and improvements
GPT‑5.3‑Codex set industry-leading benchmarks:

The model also delivered outstanding results on OSWorld and GDPval, and significant improvements were made in web development.

Versatility for all professionals
OpenAI positions GPT‑5.3‑Codex as a universal agent suitable for:

* programmers,
* designers,
* project managers,
* data specialists,
* and many other roles.

The model enables managing the entire software life cycle – from development to support.

Cybersecurity enhancement
The latest model is the first in the lineup trained directly to search for vulnerabilities in code.

As a result, the company announced an expansion of its cybersecurity grant program: $10 million in API credits to support open source software.

Availability
* GPT‑5.3‑Codex is already available to paid ChatGPT subscribers.
* Soon there will be an option to connect the model via API for third‑party developers.