OpenAI introduces GPT-5.3-Codex, a coding agent

OpenAI has introduced GPT-5.3-Codex, an agent that does more than write and review code: it can also research, use tools, operate a computer and complete complex tasks from start to finish. The company says it is its most advanced agentic coding model to date and that it runs 25% faster than before.

The difference is in how it works. You can give it a long task, such as building an application, analyzing data or preparing a presentation, and keep talking to it while it works. The agent shares its decisions and progress, takes corrections and maintains context without forcing you to start over.

From writing code to completing work

GPT-5.3-Codex combines the programming improvements in GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2. According to OpenAI, this allows it to handle more parts of a team’s daily work:

• Debugging and deploying software.
• Creating tests, metrics and documentation.
• Drafting product requirements and editing text.
• Researching users and analyzing data.
• Preparing presentations and spreadsheets.
• Using applications from a visual desktop environment.

The idea is that you do not have to turn every task into a technical request. If you ask it for a simple website, for example, it now tends to make more decisions on its own: it might show the annual price as a discounted monthly rate or add a testimonial carousel with several opinions, instead of delivering a basic mockup that you have to complete manually.

OpenAI also tested the model with web games. In one experiment, GPT-5.3-Codex created and improved games over several days, following instructions such as “fix the bug” or “improve the game.” The agent carried out autonomous iterations using millions of tokens, a unit used to measure how much text a model processes.

Better results on agent evaluations

The company says the model reaches new highs in several coding and computer-use evaluations. Its published results include:

• 56.8% on SWE-Bench Pro, a real-world software engineering test covering four languages and designed to prevent models from memorizing the answers.
• 77.3% on Terminal-Bench 2.0, which measures the ability to work from a terminal.
• 64.7% on OSWorld-Verified, a test in which the agent must complete tasks inside a visual desktop.
• 70.9% wins or draws on GDPval, an evaluation of professional tasks across 44 occupations.

On Terminal-Bench 2.0, the improvement over GPT-5.2-Codex is especially large: it rises from 64% to 77.3%. OpenAI adds that it achieves these results using fewer tokens than its previous models, although the figures come from controlled evaluations and do not guarantee the same performance in every project.

A model that helped build itself

OpenAI says early versions of GPT-5.3-Codex took part in their own development. The team used them to detect training errors, analyze results, improve deployment and identify problems in evaluations.

During testing, the model also helped identify rendering failures, investigate a low data reuse rate and create tools for analyzing thousands of records. In one case, it generated simple classifiers to measure how many clarification questions it asked, how much progress it made per turn and how users responded.

This does not mean the model was designed without human involvement. It means OpenAI used it as an active tool throughout the research and development process, which can speed up the creation of later versions.

More capability also means more controls

Cybersecurity is one of the most sensitive areas. OpenAI classifies GPT-5.3-Codex as its first model with High capability for cybersecurity-related tasks and says it trained the model directly to identify vulnerabilities.

The company says it still has no definitive evidence that the model can automate complete attacks, but it has applied additional controls because of the dual-use nature of these tools. These include automated monitoring, restricted access to advanced capabilities, security training and threat response systems.

OpenAI is also announcing $10 million in API credits for defensive research, especially in open-source projects and critical infrastructure. It is also expanding the private beta of Aardvark, its security research agent, and offering support to analyze widely used projects such as Next.js.

Availability

GPT-5.3-Codex is now available to paid ChatGPT plans in the app, command line, IDE extensions and on the web. OpenAI is working to enable API access later.

For you, the main change is not that the model writes code faster. It is that it can take on an entire chain of work: understanding a goal, researching, building, testing, fixing and explaining what it did. The important question now is how much of that autonomy holds up in real projects, where requirements change, errors have consequences and human oversight is still necessary.