Descriptions:
On April 16, 2026, OpenAI shipped a major update to Codex that transformed it from a developer command-line tool into a full Mac desktop agent capable of operating any application through computer use — clicking, typing, scheduling background tasks, and running multiple parallel agents without disrupting the user’s active window. Nate B. Jones spent the following week running Codex and Claude side by side on identical workflows and shares detailed, reproducible observations in this analysis.
The headline finding is a substantial performance gap: Codex completes computer-use workflows roughly 2–3x faster than Claude’s implementation (2 minutes versus 5–6 minutes on comparable tasks), handles a broader range of native Mac applications beyond Chrome, and rarely fails mid-task on unexpected modal dialogues — the failure mode that has made computer-use products impractical in production for the past year. Jones attributes part of the gap to GPT-5.4, which benchmarks in the mid-70s on OS World (above the human baseline for GUI control), and part to a 12-person team OpenAI acquired in October 2025: the founders of Workflow (Apple’s Shortcuts precursor) and Sky, an unreleased Mac AI interface, bringing a decade of Apple-side OS and Safari engineering to the cursor motion implementation.
Beyond benchmarks, the video frames a deeper strategic divergence: OpenAI is building an implicit, outcome-driven product where the agent selects its own interface (GUI, API, plugin, or code), while Anthropic is building an explicit, permissioned product that asks users to define scope upfront. Jones argues this reflects two genuinely different bets about how enterprise AI adoption will unfold.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published April 23, 2026
🏷️ Format: Review
