GPT Image 2.1 · Not yet announced · Status as of 3 September 2026

GPT Image 2.1 — Release Status, What We Know, and How to Try GPT Image 2 Free

OpenAI has not announced GPT Image 2.1. The current model is GPT Image 2, released 21 April 2026 and still ranked #1 on both Arena image leaderboards. This page tracks the 2.1 release, lays out what GPT Image 2 already does, and gives you a free generator running GPT Image 2 right now — no ChatGPT subscription needed.

Capabilities

What GPT Image 2 already does — the bar 2.1 has to clear

Until OpenAI ships 2.1, GPT Image 2 is the model to measure against. Six things it does today, all available in the generator above.

Thinks before it renders

GPT Image 2 is OpenAI's first image model with a reasoning step built in. It plans layout and composition before generating, so dense posters, slides, and charts arrive structured instead of stitched together.

Text that reads, in more languages

Text rendering is the headline gain over 1.5: small captions, dense paragraphs, and non-Latin scripts all hold up. OpenAI expanded language support to Japanese, Korean, Chinese, Hindi, and Bengali.

Up to eight coherent images per prompt

One prompt can return a set of up to eight images that keep the same character, product, or style across the batch — a storyboard or a carousel in a single pass.

1K, 2K, and 4K output

Standard output goes to 2K, with 4K available for print and large-format work. In the generator on this page you pick 1K, 2K, or 4K for each generation.

Up to 16 reference images

Edits accept as many as sixteen reference images, so a product, a face, or a brand palette can be held steady across a whole campaign rather than re-rolled until it matches.

Transparent backgrounds

Since 20 August 2026 the API can return a subject on a transparent background as PNG or WebP — a cutout straight from the model, with no separate masking pass.

Version History

GPT Image 1.5 vs GPT Image 2

GPT Image 2.1 has not been announced, so the most useful comparison is the last version jump. Here is what changed from 1.5 to 2 — the baseline any 2.1 release would build on.

CapabilityGPT Image 1.5GPT Image 2
Released16 December 202521 April 2026 in the API and Codex; 22 April 2026 in ChatGPT
Reasoning before renderNoneBuilt in — plans the layout, can search the web for references, and checks its own output
Max resolutionThree fixed sizes: 1024×1024, 1024×1536, 1536×1024Up to 2K as standard, 4K available (long edge up to 3840 px)
Aspect ratiosSquare, portrait, and landscape onlyFrom 3:1 ultra-wide to 1:3 ultra-tall
Images per promptOneUp to eight, with consistent characters and objects across the set
Reference imagesSupported for editsUp to 16 per edit
Text renderingImproved over GPT Image 1, but small type and non-Latin scripts still slippedReliable on small type and dense paragraphs; Japanese, Korean, Chinese, Hindi, and Bengali added
StatusNo longer the default; API removal scheduled for 1 December 2026Current model — running in the generator on this page

Specifications as published by OpenAI and its API documentation as of 3 September 2026. GPT Image is a product of OpenAI. Images 2.0 is an independent service and is not affiliated with OpenAI.

Use Cases

Where GPT Image 2 earns its keep today

A reasoning image model matters most where text, layout, and consistency all have to hold at once. Six jobs GPT Image 2 handles now — and 2.1 will inherit.

Posters & ad creative

Headline, body copy, and the legal line rendered legibly in one pass, at the sizes they will actually be read at. Fewer round trips into a design tool just to fix the type.

Product & listing images

Sixteen reference images keep a product recognisable across every angle and background a marketplace listing needs, and transparent export drops it straight onto a template.

Localised marketing assets

One layout, many languages. Multilingual text rendering means the Japanese, Korean, or Hindi version of a banner comes out of the same prompt, not out of a design tool.

Thumbnails & channel art

Short overlay text that stays sharp at thumbnail scale, with a reference image to keep a channel's look consistent from episode to episode.

Slides, infographics & diagrams

The case the reasoning step was built for: several columns, many labels, and type at a range of sizes that all has to survive in one frame.

Character-consistent series

Up to eight images from one prompt with the same character and props — a comic strip, a storyboard, or a carousel that reads as one set.

GPT Image 2.1 FAQ

Straight answers on the release, and on the model you can use today