GPT इमेज 1 मिनी

A cost-optimized image generation model that is 80% less expensive than gpt-image-1. Significantly improves text rendering, producing cleaner, more legible text even in small fonts. Accepts text and image inputs and produces image outputs.

GPT Image 1 Mini arrived without a launch post of its own. OpenAI announced it at DevDay 2025 as a single line on the developer page - a smaller image generation model 80% less expensive than the large one - alongside a matching cheap voice model, and left it at that. There are no benchmark tables and no capability claims to weigh, because the lab did not make any: the pitch is the price, and the model is the GPT Image 1 family's pipeline served at a fraction of the cost.

That makes the useful question a workflow question rather than a quality one. OpenAI's own guidance is to reach for this model when cost and throughput are the binding constraint - large batches of variants, rapid ideation, previews, lightweight personalisation, and draft assets that do not need the strongest generation or editing performance. It is explicit that this is not the model for work where a retry is expensive, and that the newer flagship at its low quality setting covers many of the same cases just as well.

The API surface is the full one rather than a cut-down version, which is what makes it usable as a drafting tier. It generates and edits, accepts a mask for inpainting, takes reference images, offers the same low, medium and high quality settings and the same three output shapes plus automatic sizing, supports transparent backgrounds, and returns PNG, JPEG or WebP. One difference is worth planning around: only a single reference image is carried at high input fidelity, against five on the newer models, so a composition that depends on several sources being preserved faithfully belongs on a larger model.

The honest way to use it is as the first pass. Explore layout, framing and direction here at a few cents an image, then re-render the two or three that survive on a model that renders text and fine detail properly. Used that way the 80% saving is real; used as a substitute for the flagship on final assets, it is a false economy that shows up in the small text.