The influence of artificial intelligence on software development is already profound, transforming the programming landscape by delegating much of the routine labor to groups of specialized AI agents. As innovators explore novel interfaces and collaborative models between humans and AI, even leading organizations in the field struggle to maintain pace with the rapid evolution. A dominant movement today is toward agentic software development—frameworks where AI can autonomously handle coding duties—exemplified by tools like Claude Code and Cowork.

Concurrently, OpenAI has been steadily advancing its own coding solution, Codex. Initially released as a command-line utility in April, it gained a web-based interface the following month. Now, the company is making a significant move to align with current trends. This week, OpenAI introduced a new MacOS application for Codex, incorporating many of the agentic methodologies that have gained traction over the past year. Designed to operate with multiple agents simultaneously, the app integrates advanced agent capabilities and modern workflows. This release follows shortly after the debut of GPT-5.2-Codex, OpenAI's most advanced coding model, which aims to attract users from competing platforms such as Claude Code.

"If you truly aim to tackle sophisticated and complex projects, 5.2 is by far the most capable model available," stated CEO Sam Altman during a briefing. "However, its complexity has made it challenging to use. Embedding that level of power within a more adaptable interface, we believe, will make a substantial difference."

Although Altman's assurance in GPT-5.2 is warranted, performance metrics present a nuanced picture. The model currently leads on TerminalBench, an assessment focused on AI proficiency with command-line programming tasks, at least at the time of reporting. Yet, agents based on Gemini 3 and Claude Opus have recorded comparable results—slightly lower but within the benchmark's margin of error. Findings from SWE-bench, another evaluation that measures an AI's ability to resolve actual software bugs, are alike, showing no definitive superiority for GPT-5.2. Benchmarking agentic applications remains inherently difficult, and leading models can differ greatly in real-world usability.

The new Codex application includes a suite of features intended to help it match or even surpass various Claude applications. It supports automated processes that can execute in the background on a preset schedule, with outcomes queued for review upon the user's return. Additionally, users can customize their agent's demeanor—from practical to understanding—to suit their personal workflow.

For OpenAI, however, the primary advantage is the remarkable acceleration in development that AI enables. "Starting from a blank slate, you can create a highly sophisticated software application in just a few hours," Altman explained. "The speed at which I can input new ideas is the only constraint on what can be constructed."