# Image generation adds GPT Image 2.5 and Nano Banana 2.1

> Workflows and apps can now choose the image model, edit reference images, and draw at up to 4K, with Nano Banana 2.1 as the new default.

Oct 21, 2026. Topics: Models, Workflows, Developers. https://www.generalinput.com/changelog/2026-10-21-image-generation-model-choice

The image generation service your workflows, apps, and specialists call now lets you choose the model. Nano Banana 2.1 from Google replaces Gemini 2.5 Flash Image as the default: it renders text inside images more legibly and draws at up to 4K. GPT Image 2.5 Sunburst from OpenAI is the choice when quality matters most, or when an existing image needs a precise edit. Both now take reference images, so an agent can restyle a product photo, put a logo on a mockup, or keep a character the same across a series. The [developer docs](/developers) cover the request format.

### How it works

- **Name a model, or don't.** A request that leaves out the model gets Nano Banana 2.1, so existing workflows keep working without changes and move to the new default.
- **Reference images.** Up to 14 on Nano Banana 2.1 and 16 on GPT Image 2.5, sent as the image bytes themselves.
- **Sizes and quality.** Both models draw seven aspect ratios, from square to 16:9 and 9:16. Nano Banana 2.1 draws at 1K, 2K, or 4K. GPT Image 2.5 draws at 1K, with low, medium, or high quality.
- **Billed by what each image uses.** Each image is billed by the tokens its model reports, so a 4K image or a high quality one costs more than a quick 1K draft.
- **Clear refusals.** A setting the chosen model cannot take is refused before anything is drawn, and a model that declines a prompt says so instead of failing silently.