OpenAI launches GPT-5.2-Codex for coding

OpenAI has launched GPT-5.2-Codex, an AI model designed to work for longer on complex software projects. It is already available to paid ChatGPT users through Codex tools, while API access will arrive in the coming weeks.

A model for long-running projects

GPT-5.2-Codex is built on GPT-5.2 and optimized for agent-assisted coding. That means it does more than suggest a line of code: it can analyze a repository, use terminal tools, modify multiple files and check whether its changes work.

Its main improvement is maintaining context during extended sessions. If the project is large, the model can summarize and reorganize information it has already processed without losing sight of the goal. In practice, this helps with tasks such as:

• Reorganizing large parts of an application.
• Migrating code from one technology to another.
• Building features that affect many files.
• Continuing work even when initial attempts fail or the plan changes.

OpenAI also says the model is more reliable when calling tools, better at understanding long instructions and able to produce more precise responses while using fewer tokens in its reasoning.

Improvements in Windows and image understanding

GPT-5.2-Codex improves work in Windows environments, one of the areas where its predecessor, GPT-5.1-Codex-Max, had already made progress. This matters because a tool that works well on Linux does not always behave the same way when handling Windows-specific commands, paths and configurations.

The model also interprets screenshots, technical diagrams, charts and interface designs more effectively. You can show it a visual mockup and ask it to turn that design into a working prototype. That prototype can then continue evolving into a production-ready application, although human review is still necessary.

According to OpenAI, GPT-5.2-Codex achieves benchmark results on SWE-Bench Pro and Terminal-Bench 2.0, two tests that evaluate how agents solve coding tasks in real repositories and terminals.

More capability to find vulnerabilities

The most sensitive change is in cybersecurity. OpenAI says GPT-5.2-Codex has greater capabilities than any previous model from the company for investigating and analyzing security flaws.

The defensive value is clear: a team can use it to review code, reproduce a bug, test possible attacks and speed up the creation of a fix. The risk is just as clear. Those same capabilities could help someone search for vulnerabilities with malicious intent.

OpenAI cites a recent case involving React. Andrew MacPherson, a security engineer at Privy, used GPT-5.1-Codex-Max with Codex CLI to study a vulnerability known as React2Shell. During that process, the system helped prepare a testing environment, analyze attack surfaces and run tests with altered inputs. The work ultimately uncovered other vulnerabilities, which were responsibly disclosed to the React team.

Gradual access to reduce risks

OpenAI says GPT-5.2-Codex has not yet reached the “High” level of cyber capability defined in its Preparedness Framework. Even so, the company says it is preparing as though upcoming models could cross that threshold.

That is why the launch includes more controls in both the model and the product. OpenAI is also preparing a restricted-access program for pre-screened security professionals and organizations. At first, it will be invite-only and aimed at people with demonstrated experience in responsible vulnerability research, authorized testing and digital defense.

For you, the most visible change will be a tool capable of taking on broader, longer-running coding tasks, rather than simply answering questions about code. For companies, the question will be how much work to delegate and how to review each change before incorporating it into real systems.

GPT-5.2-Codex also marks a new stage in the race between digital defense and attack. OpenAI is expanding access gradually because it expects models to keep improving rapidly in cybersecurity. What matters now is whether safeguards grow at the same pace as capabilities.