What is GPT-5.3-Codex, and what does it mean for a coding model to work beyond code?

Yesterday, February 5, 2026, OpenAI introduced GPT-5.3-Codex, a new model behind Codex, its coding agent. It arrived three days after the new Codex app for Mac, and, as ZDNET pointed out, at the very moment Anthropic was releasing a major model of its own.

The name tells you what OpenAI merged. Until now there were two lines: GPT-5.2-Codex, tuned for writing and reviewing code, and GPT-5.2, the general model with stronger reasoning and professional knowledge. GPT-5.3-Codex puts both into one model and runs 25% faster for Codex users. OpenAI's goal is an agent that does not stop at code, but handles most of what a developer or other professional does at a computer.

What OpenAI released

The model is available now on paid ChatGPT plans wherever Codex runs: the app, the command-line tool, the editor extension and the web. API access, for developers building their own products, is promised "soon." Free users in the current Codex promotion stay on GPT-5.2-Codex for now, according to ZDNET.

OpenAI published these scores, all at the highest reasoning setting:

The biggest jumps are in the terminal and on the desktop. On OSWorld, OpenAI notes that people score about 72%, so the model is getting close to human level at clicking through ordinary software. OpenAI adds that it reaches its coding results using fewer tokens, the small text pieces that models read and write, than earlier models, which means lower cost for the same work.

How it works: from coder to computer worker

Three changes matter most for everyday use.

It goes beyond code. OpenAI lists debugging, deploying, monitoring, writing product requirement documents (PRDs, the specs that describe what to build), editing text, user research, tests and metrics. It also shows the model making slide decks and spreadsheets, such as a financial-advice presentation and a net-present-value analysis.

It guesses intent better. Given a short, vague request for a website, it now fills in sensible features by default. In OpenAI's example, a pricing page showed a yearly plan as a discounted monthly price and added a rotating testimonials section with three quotes, where the previous model did less.

You can steer it mid-task. Instead of waiting for a finished result, you get frequent updates and can reply while it works: ask why it chose an approach, or redirect it. OpenAI compares this to working with a colleague. In the Codex app you turn it on under Settings, General, Follow-up behavior.

OpenAI also tested long, autonomous work. Using generic follow-ups like "fix the bug" or "improve the game," the model kept improving a racing game and a diving game on its own over millions of tokens. ZDNET reports Codex can now run jobs lasting more than a day.

A model that helped build itself

OpenAI calls GPT-5.3-Codex its first model that was "instrumental in creating itself." Early versions helped the team watch and debug the training run, find bugs in how context was displayed to users, adjust computing capacity during launch traffic, and analyze test data. In one case, a researcher asked it to measure how much more work it completed per turn; it wrote simple pattern-matching classifiers, ran them over the session logs and produced a report.

This is not a model training itself unsupervised. People ran the project; the model acted as a fast, capable assistant inside it. ZDNET's reviewer admitted he wasn't sure whether to find that exciting or unsettling.

The cybersecurity question

Under its Preparedness Framework, OpenAI's system for rating serious risks, this is the first model classified as High capability in cybersecurity, and the first trained directly to find software vulnerabilities. OpenAI says it has no firm evidence the model can carry out attacks end to end, but is acting with caution:

• safety training and automated monitoring;
• Trusted Access for Cyber, a pilot giving vetted defenders access to advanced abilities;
• a wider private beta of Aardvark, its security research agent;
• $10 million in API credits for defensive work, especially on open-source and critical infrastructure.

The model was co-designed with, trained on and served from NVIDIA GB200 NVL72 systems.

How do you steer GPT-5.3-Codex while it works?

First enable steering in the Codex app's settings, under Follow-up behavior. Then start with a task that has a clear goal and a few limits, and plan to talk during the work rather than only at the end:

Add a search box to the product list page. Filter as the user types.
Post a short update after each step. Before changing any shared
component, stop and explain your plan so I can approve it.

As updates arrive, reply in plain language: "Use the existing button style," "Skip the animation," or "Why did you choose that library?" The model keeps its context, so you don't need to restart.

Keep a few habits. Work in a Git branch so every change can be undone. Ask for a summary of what changed and how to test it. And for anything touching security, review the code yourself: the same skill that finds vulnerabilities makes careful human checking more important, not less.