OpenAI launches a new macOS application for agentic coding

AI has already begun to affect the way software is written, with a large amount of programming work now being done by teams of agents and subagents. But as developers try new forms and shapes of human interaction with AI, it becomes difficult for even the most advanced AI labs to achieve.

The current trend is in software development – systems where AI agents can work independently on scripts – captured by Claude Code and Cowork apps. Meanwhile, OpenAI has been slowly building its Codex toolkit, which it introduced as command line tool the month of April is extended to the web interface after a month.

Now OpenAI is taking a big step towards it. On Monday, the company launched a new initiative software for macOS to Codex, incorporating many medical practices that have become popular over the past year. The new program is designed to work with multiple agents in sync, including technical support and other operational technologies. The installation comes back less than two months later implementation of GPT-5.2-CodexA very powerful example of OpenAI, which the company hopes will be enough to test users of Claude Code.

“If you really want to do the most demanding tasks, the 5.2 is the most powerful model,” CEO Sam Altman told reporters in a press conference. “However, it has been difficult to use, so taking part of the model and putting it in a flexible environment, we think will be very important.”

While Altman’s confidence in GPT-5.2 is understandable, the benchmarks tell a more complicated story. GPT-5.2 works top spot on TerminalBench (a test to test the performance of AI for command-line programming), especially at print time. But the helpers from Gemini 3 and Claude Opus came in with similar scores – low, but within the margin of error of the benchmark. Results from SWE-benchanother benchmark that tests the ability of AI to correct global software errors, is the same, which does not show the obvious advantage of GPT-5.2. However, use cases for agents have been difficult to label well, and current models can vary widely in terms of use.

The Codex software also comes with a number of new features that OpenAI says will make it as efficient as or, in some cases, better than Claude’s various software. Codex software will allow automations to be set up to run in the background on an automatic schedule, with the results queued up for review when the user returns. Users can also choose different personalities of the agent – from pragmatic to empathetic – depending on their work environment.

But for the company, the biggest selling point is the speed of development made possible by AI. “You can use this from a clean, fresh sheet of paper to create a high-quality program in a matter of hours,” Altman said. “The more I can write new ideas, that’s the limit of what can be built.”