GPT Image 2: reassess output sizes before changing the default

The model release

OpenAI released gpt-image-2 on April 21, 2026, expanding the GPT Image family.

Treat dimensions as an application contract

A generated image is usually consumed by another system: a content manager, ad template, product page, or editor. List the actual dimensions and file requirements of those destinations before choosing output settings.

When a new model offers additional choices, resist exposing all of them immediately. Start with the presets your downstream workflow can handle, and check the current model reference for supported values.

Inspect the whole delivery path

Generate a candidate, save it, prepare the required crops, and load it into the final template. Check important details near the edges as well as the center. An image can pass an isolated visual review and still fail a placement with a different aspect ratio.

The architectural concept below is illustrative. It suggests why framing matters, but it is not evidence of a particular model capability.

Update cost estimates with the presets

Compare candidate count, output settings, and accepted results together. Keep old and new request configurations in the evaluation record. Once the rollout is complete, remove unused presets so the interface continues to reflect the workflow your team actually supports.