OpenAI launches GPT-5.3-Codex: 25% faster and for more than code

What is new

On February 5, 2026, OpenAI introduced GPT-5.3-Codex, the latest version of the company's Codex model. Codex started as a tool for writing code but is now also meant as an agent that can use a computer and take care of a broader range of work tasks. An agent is an AI program that can carry out a series of steps on its own, one after another, instead of just answering a single question.

The model is up to 25 percent faster than GPT-5.2-Codex. It can work on long runs: find bugs, make fixes in several places in a program, write and run tests and manage the deployment of a new version. You can give it input along the way so it stays on course.

OpenAI also says GPT-5.3-Codex is the first of the company's models that helped create itself. Early versions were used to debug the training, manage the deployment and interpret test results.

What is clever

The clever part is that the model combines coding abilities with better general understanding and professional knowledge in one system. You do not need to switch between different models for different types of tasks.

Tasks that previously required many small questions and answers can now run as one long workflow where you can step in and steer along the way.

The measurements show a clear jump in some areas. On Terminal-Bench 2.0, which tests work in a command window, the model scores 77.3 percent versus 64.0 for GPT-5.2. On OSWorld-Verified, which tests whether AI can use an ordinary computer, the figure is 64.7 percent versus 38.2. On SWE-Bench Pro, a test of real programming tasks, the difference is small: 56.8 versus 56.4 percent.

Cheaper or better?

Better, but not cheaper. According to independent sources, the API price is unchanged from GPT-5.2-Codex: $1.75 per million tokens (pieces of text) for what you send in and $14 per million tokens for the answers. If input is reused from a cache, it costs $0.175 per million tokens.

For ordinary users, there is no extra cost if you already have a paid ChatGPT plan. The monthly price for Codex is not stated in the material I have found.

I have not found a direct comparison with competitors such as Anthropic and Google in the sources, so I cannot give one here.

What it is good at

The model suits programmers and IT teams who want an AI to take care of longer tasks: fixing bugs, building new features, testing and shipping new code. The big improvement in computer use also makes it interesting for companies that want to try AI for office work other than coding.

GPT-5.3-Codex is available on paid ChatGPT plans everywhere Codex is offered: in the Codex app, in the command line, as an extension in programming tools and on the web. It has also rolled out in GitHub Copilot and Microsoft Foundry, and OpenAI has opened API access for developers. Whether there are special terms for Denmark or the EU does not appear in the sources.