GPT Image 2 and 2.5 — OpenAI's image model

Top view of a creative agency desk: poster sketches, a tablet with a drawing, markers, stickers and coffee

Key facts

Developer	OpenAI
Versions	GPT Image 2 (April 21, 2026), GPT Image 2.5 Flare and Sunburst (September 8, 2026)
What it does	Generates and edits images, including by mask and from several references
Resolution	Up to 3840 px on a side and 8.3 MP, aspect ratio from 1:3 to 3:1
Quality (2.5)	low, medium, high, xhigh, max or auto
Background	Transparent — in PNG and WebP
Where it runs	ChatGPT, Codex, OpenAI API
In AI2Face	Yes

What is GPT Image

GPT Image is OpenAI’s family of image models. They run inside ChatGPT (when you ask it to draw) and are available to developers through the API. In May 2026 OpenAI removed the older DALL·E 2 and DALL·E 3 from the API — pictures in the OpenAI ecosystem are now made only by GPT Image.

GPT Image 2 launched on April 21, 2026, together with ChatGPT Images 2.0. OpenAI focused on hard visual tasks: layout, diagrams, infographics, comics and accurate text in many languages. Paired with reasoning models, generation can look up information, make variants and check the result on its own.

GPT Image 2.5 launched on September 8, 2026 — ChatGPT Images 2.5 and two API models. Flare is the main, fast one. Sunburst is more accurate on edits, but slower. OpenAI promises sharper detail, more natural light and textures, better likeness to references and latency up to 50% lower than 2.0.

What GPT Image can do

Text and layout. The model’s main strength is lettering you do not have to fix: lines in a comic, menus, posters, interfaces, diagrams. Cyrillic works too — the comic and the menu in the gallery were made on the first try.

Interfaces and mockups. App screens, landing pages and presentations with readable labels — useful for prototypes and pitches.

Edits. The model edits uploaded images by mask (inpainting) or by text alone. In 2.5, multi-step edits became more reliable. ChatGPT added comments directly on the picture and sketching as a hint.

Several references. You can pass several images and assemble one scene from them — for example, a gift set from photos of separate products. OpenAI does not state an exact limit.

Transparent background. GPT Image outputs PNG or WebP with no background: stickers, icons, cut-out products for listings.

Quality for the job. In 2.5 there are six quality levels — from low for drafts to max for the final — and custom sizes up to 3840 px.

GPT Image 2 and 2.5: how they differ

	GPT Image 2	GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Released	Apr 21, 2026	Sep 8, 2026	Sep 8, 2026
For	general use	default, faster	detailed creative work
Speed	baseline	latency up to 50% below 2.0	slower than Flare
Multi-step edits	yes	more reliable	the most accurate
Likeness to references	good	better	better
Quality levels	low, medium, high	low → max, auto	low → max, auto

A ballet dancer ties the ribbons of her pointe shoes backstage: GPT Image 2 on the left, the same prompt in GPT Image 2.5 on the right — One prompt in two versions: GPT Image 2 on the left, GPT Image 2.5 Flare on the right — softer light, richer texture in the fabric and the floor

What GPT Image is good for

Posters, menus, covers with a lot of text.
Comics and storyboards with the same characters in every panel.
Interface prototypes and slides for presentations.
Infographics and diagrams.
Products on a transparent background for marketplace listings.

GPT Image and its alternatives

	GPT Image 2.5	Nano Banana Pro	Nano Banana 2	Seedream 5.0 Pro
Developer	OpenAI	Google DeepMind	Google DeepMind	ByteDance Seed
Released	Sep 8, 2026	Nov 20, 2025	Feb 26, 2026	Jul 8, 2026
Maximum size	3840 px on a side	4K	4K	2K
Text	strong, many languages	strong, translates lettering	good	10+ languages, layers
Transparent background	yes	no	no	no
Web search	via reasoning models	yes	yes	no
In AI2Face	yes	yes	yes	yes

GPT Image 2.5 is the most convenient model for text, interfaces and a transparent background. Nano Banana Pro is stronger on complex scenes and face likeness from several photos, Nano Banana 2 is faster for everyday tasks, and Seedream 5.0 Pro is for layout with layers and precise edits.

How to write prompts for GPT Image

Put text in quotes, each element on its own: the headline, the lines, the captions. Add “all text exactly as given, no other text”.
For comics, number the panels and describe the action and the line in each one.
For interfaces, list the on-screen elements with exact values: «Steps 7,842 / 10,000».
For photos, write like a photographer: lens, light, film, skin texture.
For drafts, use low quality; for the final, use high or max.

How to use GPT Image in AI2Face

GPT Image is available in the AI2Face generator: choose the model, then add a prompt or references. The cost is shown before you start.

For the same jobs, the Nano Banana family and Seedream sit next to it:

Nano Banana Pro — posters and infographics with accurate text, up to 4K, up to 14 references.
Seedream 5.0 Pro — complex layout, text in 10+ languages, edits by coordinates.
Nano Banana 2 — a fast general-purpose model for everyday pictures.

All three are available in the AI2Face generator, and the cost is shown before generation. Prompts from the gallery on this page work for them too.

Limitations

Text is not perfect. OpenAI says the model can still miss the exact placement and sharpness of text — check small lettering.
Repeated elements. Keeping identical objects consistent across a series is still a hard task.
Speed. Complex prompts in the API can take up to 2 minutes to generate.
Moderation. Every request and picture is filtered by OpenAI’s rules; blocked topics return an error.
Access. Officially it’s ChatGPT, Codex and the OpenAI API.

Conclusions

GPT Image 2.5 is the best choice when a picture carries a lot of text or looks like an interface: menus, comics, posters, prototypes, products on a transparent background. Flare fits most jobs, Sunburst fits detailed work with edits. In AI2Face, GPT Image is available alongside Nano Banana Pro and Seedream: prompts from this page work on all three, while for photorealism and complex scenes with people, Pro and Seedream win more often.

Generation examples

Four-panel comic about a cat and a robot vacuum, with lines in Russian — **Prompt**
A four-panel comic in a clean modern style about a cat and a robot vacuum. Panel 1: the cat looks at the vacuum, line «Что это за зверь?». Panel 2: the robot bumps into the cat, sound «БУМ!». Panel 3: the cat rides the robot, line «Такси подано». Panel 4: the cat sleeps on the charging dock, caption «Новый лучший друг». The same characters in every panel, bold outlines, pastel colors, all Russian text exactly as given

Breakfast menu on cream paper: five dishes with prices in rubles and opening hours — **Prompt**
An elegant café menu in Russian. Title «МЕНЮ ЗАВТРАКОВ». Five items with prices in rubles, each on its own line: «Сырники со сметаной — 390 ₽», «Омлет с травами — 350 ₽», «Каша с ягодами — 290 ₽», «Тост с авокадо — 420 ₽», «Капучино — 210 ₽». A small line at the bottom «Завтраки с 8:00 до 12:00». Cream paper texture, dark green serif type, line drawings of a cup and a whisk, all text exactly as given, no other text

Fitness app screen on a phone: steps, heart rate, sleep and water, lime accent on a dark theme — **Prompt**
A fitness app screen on a smartphone: greeting «Good morning, Anna», a ring chart «Steps 7,842 / 10,000», cards «Heart rate 72 bpm», «Sleep 7h 20m», «Water 1.2 L», a bottom navigation bar with four icons. Clean modern design, a soft lime accent on a dark theme, all text sharp and error-free, no brand logos

An older vendor with a white beard arranging oranges at a morning market — **Prompt**
A candid street photo of a 60-year-old man with a white beard selling oranges at a market stall in a southern European town in the morning: wooden crates, handwritten price cards with no readable text, warm low sun, 35 mm film, natural skin texture, photorealism

FAQ

What is GPT Image?

OpenAI's image generation models, built into ChatGPT and available through the API. The current generation is GPT Image 2.5 in two versions, Flare and Sunburst; before that, GPT Image 2.

How is GPT Image 2.5 different from GPT Image 2?

According to OpenAI, 2.5 gives sharper detail, more natural light and textures, keeps characters from references better and follows multi-step edits more closely. Latency is up to 50% lower than 2.0.

What are Sunburst and Flare?

Two versions of GPT Image 2.5 in the API. Flare is the default: quality above GPT Image 2 at half the latency. Sunburst is for detailed creative work: more accurate on edits, but slower to generate. In ChatGPT you cannot pick the version.

Does GPT Image render text in other languages?

Yes, and that is one of the model's strengths. In our tests GPT Image 2.5 wrote the comic lines and a five-item menu with prices without errors. OpenAI still warns that exact placement and sharpness of text sometimes give the model trouble.

What is the maximum resolution of GPT Image?

Up to 3840 pixels on the long side and no more than 8.3 megapixels, with an aspect ratio from 1:3 to 3:1. The “4096 × 4096” figure that shows up in reviews is not confirmed by OpenAI's documentation.

Is GPT Image in AI2Face?

Yes. GPT Image is available in the AI2Face generator, and the cost is shown before generation. For the same jobs you also have Nano Banana Pro, Seedream 5.0 Pro and Nano Banana 2 there.

Sources

https://openai.com/index/introducing-chatgpt-images-2-0/
https://openai.com/index/introducing-chatgpt-images-2-5/
https://developers.openai.com/api/docs/models/gpt-image-2
https://developers.openai.com/api/docs/models/gpt-image-2.5-flare
https://developers.openai.com/api/docs/models/gpt-image-2.5-sunburst
https://developers.openai.com/api/docs/guides/image-generation
https://developers.openai.com/api/docs/changelog