OpenAI’s Codex Mac app takes aim at Claude Code and Gemini agents

OpenAI

• OpenAI launched a dedicated macOS app for Codex with agent-driven workflows
• The app supports parallel agents, shared context, and flexible state management
• Codex excels in some tasks, but Claude and Gemini are close competitors.

Did our AI summary help?

OpenAI launched a dedicated macOS app for Codex, folding in many of the agent-driven workflows that have gained traction across the industry. The new app supports multiple agents running in parallel, shared context between tasks, and more flexible state management, all designed to mirror how developers actually work on complex projects.

The release follows the debut of GPT-5.2-Codex less than two months ago, OpenAI’s most capable coding model to date. The company is betting that pairing this model with a more intuitive interface will help lure developers who have gravitated towards Claude’s tooling. Speaking to reporters, CEO Sam Altman said the challenge was never raw capability, but usability. According to him, the new app lowers the friction that previously made Codex feel harder to work with on large, evolving codebases.

Benchmarks, however, paint a mixed picture. GPT-5.2 currently leads TerminalBench, which measures performance on command-line programming tasks. Yet rival agents from Gemini 3 and Claude Opus post scores close enough to fall within the benchmark’s margin of error. SWE-bench, which focuses on fixing real-world software bugs, shows a similarly crowded field with no decisive winner. Measuring agentic workflows remains especially difficult, and real-world experience often diverges from lab results.

Beyond core coding, the Codex app introduces several practical features. Users can schedule background automations that run while they are away, with outputs queued for later review. Agents can also be tuned with different personalities, ranging from pragmatic to empathetic, allowing developers to shape how the system communicates and prioritises tasks.

Altman argues that with the right agentic setup, a sophisticated piece of software can be built from scratch in a matter of hours. In his view, typing speed, not model capability, is becoming the primary constraint. Whether that promise is enough to shift developers away from entrenched rivals will become clearer as Codex enters everyday use.