Artificial intelligence is fundamentally reshaping software creation, as numerous agents and subagents now handle a significant portion of coding’s laborious tasks. However, the rapid pace of innovation in human-AI collaboration interfaces and forms presents a challenge even for leading AI research institutions to stay current.
Agentic software development, a paradigm where AI agents autonomously manage coding assignments, represents the prevailing trend, exemplified by applications like Claude Code and Cowork. Concurrently, OpenAI has been progressively developing its Codex tool, which initially debuted as a command line utility in April, subsequently evolving into a web-based interface the following month.
OpenAI is now making a significant move to advance its position. This Monday, the organization unveiled a new MacOS application for Codex, incorporating numerous agentic methodologies that have gained traction recently. This fresh app facilitates parallel operation with multiple agents, integrating agent skills and cutting-edge workflows. Its release follows closely, less than two months after the introduction of GPT-5.2-Codex, OpenAI’s most potent coding model, which the company anticipates will attract users from Claude Code.
“For tackling intricate and complex tasks, GPT-5.2 stands out as the most robust model,” CEO Sam Altman informed journalists during a press conference. “Nevertheless, its usability has been challenging; therefore, integrating such advanced model capabilities into a more adaptable interface, we believe, will be highly impactful.”
Although Altman’s strong belief in GPT-5.2 is justifiable, an examination of coding benchmarks reveals a more nuanced picture. GPT-5.2 currently occupies the leading position on TerminalBench, a benchmark evaluating AI proficiency in command-line programming, as of this publication. Yet, agents from Gemini 3 and Claude Opus have achieved comparable scores, slightly lower but falling within the benchmark’s margin of error. Similar findings emerge from SWE-bench, another coding benchmark assessing AI’s capability to resolve real-world software defects, indicating no distinct superiority for GPT-5.2. It’s worth noting, however, that benchmarking agentic applications has proven challenging, and user experience can differ considerably among cutting-edge models.
The Codex application also introduces several new functionalities, which OpenAI states will enable it to match or, in certain instances, surpass the capabilities of the various Claude applications. Users of the Codex app can configure automations to execute in the background according to an automatic schedule, with their outputs queued for review upon the user’s return. Additionally, users have the option to assign diverse personalities to the agent, ranging from pragmatic to empathetic, to align with their preferred working methodology.
However, for OpenAI, the most compelling advantage lies in the remarkable development velocity facilitated by AI. Altman stated, “One can employ this, starting from scratch, to craft a remarkably sophisticated software application within mere hours. The pace of new ideas I can input dictates the boundary of what can be constructed.”