Artificial intelligence is entering the era of self-improvement.
On Thursday afternoon, OpenAI released a new cutting-edge coding model that the company said assisted in its own creation.
“GPT-5.3-Codex is our first model that was instrumental in creating itself,” OpenAI stated in a blog post accompanying the model’s release. “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.”
As an AI-powered coding model, Codex allows users to build complex software programs and applications using instructions written in natural language. While AI researchers and engineers used to write individual lines of code, coding models such as Codex, Anthropic’s Claude Code and Cursor outsource line-by-line writing to AI, while humans are only required to provide high-level instructions.
These new AI systems and coding assistants are revolutionizing how developers write software, but they’re also changing how AI companies themselves are developing the next generation of AI models and coding assistants.
This tightening feedback loop, in which AI systems help design their successors, are changing the business of AI and breaking down walls between internal teams, while also raising questions about safety and the role of human engineers.
The new coding model released Thursday afternoon, entitled GPT-5.3-Codex, builds on OpenAI’s GPT-5.2-Codex model and combines insights from the AI company’s GPT-5.2 model, which excels on non-coding tasks like text analysis and reasoning.
OpenAI says the new model runs 25% faster than the previous version and achieves state-of-the-art performance while using fewer computing resources. OpenAI also said the model is the first to be designated as “high-capability” for cybersecurity-related tasks as defined by its internal risk-management assessments and guidelines.
Increasingly capable coding models have been projected to ease hackers’ or bad actors’ ability to conduct cyberattacks, while also potentially boosting cyber defenses.
Alex Embiricos, who leads the Codex team’s product division, said the increasing speed and capabilities of coding agents is supercharging AI development. “If you look at where we were and the amount of progress we’ve had in the past year on our models’ coding ability, it’s crazy,” Embiricos told NBC News. “I think the velocity there has been super high.”
“Researchers are using Codex themselves,” he said. “A researcher will do a training run, and they’ll be using Codex to ‘babysit’ or monitor that training run, or they’ll use Codex to analyze some data about the training run, or they’ll use Codex to clean up a data set or something like that.” Embiricos said developers and engineers across OpenAI are using Codex not only to write new code, but also to validate and evaluate code to ensure it is effective and efficient.
Beyond creating the underlying models that generate code, AI is also being harnessed to create the interfaces that customers and developers outside the company use to write code. On Monday, OpenAI released a Codex desktop app that is meant to assist in managing multiple largely autonomous AI computing tasks simultaneously.
The app, meant to enable users to more easily create code and orchestrate multiple autonomous personal coding assistants, or AI coding agents, has already been downloaded more than 500,000 times.
Ed Bayes, who leads product design for Codex, said the rise of Codex and coding models had completely changed how he designs products like the app. “I spend 90% of my time in code now. A year ago, it was flipped, and it was maybe 10%,” Bayes told NBC News, reminiscing about using graphic design software or laying out design prototypes using Photoshop. “It’s radically changing the way that people are able to build and prototype.”
Bayes said the new app, meant to both appeal to seasoned coding veterans and welcome less experienced coders with its intuitive layout and natural-language inputs, was itself a product of increasingly powerful Codex models. “A year ago, you couldn’t build software in such a short period of time without it looking like it was built in such a short period of time. But now, engineers are becoming designers. Designers are becoming engineers. I think these walls are starting to kind of fall down.”
AI engineers and researchers at companies such as OpenAI and rival Anthropic have long sought to apply AI systems to the problem of developing AI models themselves. In October, OpenAI CEO Sam Altman stated the company aimed to have “an automated AI research intern by September” and a “true automated AI researcher by March of 2028.”
Many experts credit AI’s improving software-writing abilities to the fact that coding tasks can be easily graded or evaluated by an AI system. If an AI system can easily confirm that a certain task is completed, that positive feedback can inform the system’s future development. This quick feedback cycle is critical to current training methods for AI systems, which are highly dependent on so-called reinforcement learning, and eases automation of research tasks.
Experts have predicted that AI systems capable of autonomously improving themselves, otherwise known as recursive self-improvement, might vastly increase the speed at which AI development takes place, potentially shortening years of complex research into weeks or even days. Some foresee this leading to a sort of explosion in the intelligence of AI systems, while others caution that such iteration might work to increase capabilities in some domains — like software engineering — while AI systems will continue to struggle on cognitive or physical abilities.
Early last week, Anthropic CEO Dario Amodei told NBC News that the self-improving phenomenon is also starting to take shape at Anthropic. “We essentially have Claude designing the next version of Claude itself, not completely, not in all ways, but in many ways, that loop starts to close very fast.”
