Last quarter I shipped a single landing hero with 47 image iterations across four tools. This week I shipped three landing pages, two onboarding flows, and a full pricing section in the same tool I write code in. The thing that broke the loop is not faster pixels, it is reasoning before pixels.

GPT Image 2 is OpenAI's April 21, 2026 image model that runs inside the same O-series reasoning loop as the rest of Codex, accepts up to 16 reference images, and renders natively at 1K, 2K, and 4K. The pinned snapshot is gpt-image-2-2026-04-21 . It ships in three places at once: ChatGPT Images 2.0 for consumers, the OpenAI API in early May, and as a first-class tool inside the Codex App and Codex CLI . Microsoft Foundry shipped it on day one too.

The headline number for me is not resolution. It is iteration count. My average dropped from 47 to 6.

The pain I built this around

For two years my frontend loop was a relay race between four runners who kept dropping the baton.…