GPT Image 2.5 and ChatGPT Images 2.5: What's New, How to Use It

ChatGPT Images 2.5 was released on September 8, 2026. It is OpenAI's new image model: in the app it is called Images 2.5, in the API it is split in two, GPT Image 2.5 Flare (fast, for everyday tasks) and GPT Image 2.5 Sunburst (slower, more precise in edits). Searches like "chatgpt images" and "chat gpt pictures" lead to it as well.

According to OpenAI, light and textures became more natural, but the main change is in control: the model edits only what you asked for, keeps faces and objects from reference photos recognizable and does not fall apart after the fifth edit in a row. Plus three new tools inside ChatGPT itself: the Sketch canvas, templates for posters and product photos, and comments right on the image. In Pixyn both 2.5 versions are open: in the studio as a regular generation chat, in actions as an advanced editor.

TL;DR

• What was released. On September 8, 2026 OpenAI updated the image generator in ChatGPT to version 2.5 and released two API models: gpt-image-2.5-flare and gpt-image-2.5-sunburst.
• What got better. Precise edits without side changes, stability under multi-step editing, better preservation of people and objects from photo references, transparent background, complex layouts. Speed went up: for the fast Flare version OpenAI claims up to 50% lower latency compared with the previous Images 2.0 (that is GPT Image 2 in the API).
• New tools in ChatGPT. Sketch (you draw a rough sketch, the model finishes it), 16 templates from a poster to a website mockup, comments on an area of the image, the option to share an image together with its prompt.
• For developers. Quality from low to max, custom sizes up to 3840 pixels per side, a mask for inpainting. The per-token rates are the same as for GPT Image 2.
• Where to use it in Pixyn. GPT Image 2.5 already works in the Pixyn studio (a regular generation chat) and in the "ChatGPT Images" action (an advanced editor with templates, sketches and versions).

GPT Image 2.5 in Pixyn: a regular chat and an advanced editor

Pixyn is connected to OpenAI through the official API, so the ChatGPT image models here respond the same way as in the original, just without the ChatGPT app itself. The 2.5 model is available in two forms, and they do not replace each other.

A regular generation chat in the studio. Open the studio and choose the ChatGPT group in the list of image models. It contains GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, GPT Image 2 and GPT Image 1. You write a prompt, choose a format (1:1, 3:2, 2:3, 16:9, 9:16) and a resolution of 1K, 2K or 4K, and get an image. Cost: from 4 tokens per image for both Flare and Sunburst. Every generation stays in the history together with the prompt and settings.

To edit your own photo, attach it to the prompt: the model switches to edits mode and reworks the uploaded image according to the instruction. Need a hard edit boundary? Upload a mask: a PNG of the same size where the transparent area is the zone to change.

An advanced editor in actions. The "ChatGPT Images" action is built as a workspace, not as a chat. Here is what it has:

• a version timeline: each edit builds on the active version, you can return to any previous one with one click and continue from it;
• references next to the prompt: style, object, sketch; together with the source the vendor takes up to 16 images at once, the prompt refers to them by number, and the source is Image 1;
• the "Sketch" canvas: pen, eraser, line, arrow, frame, oval, text, thickness, 13 colors and a custom one; on an empty canvas the drawing immediately becomes the source and the sketch-to-photo style turns on with a prompt from the OpenAI guide, and when an image is already open the sketch goes in as a layout reference;
• 16 format templates (poster, logo, product photo, infographic, book cover, website mockup and others) as short forms with two or three fields; the prompt skeleton is assembled by the OpenAI formula automatically, and your required text goes in quotes with an instruction to show it once;
• 39 styles of our editor: stickers, anime, a Tarot card, a blueprint, film frames and the rest;
• "Fast" mode (Flare) or "Best quality" (Sunburst), a transparent background for logos and icons, the same formats and resolutions as in the studio.

When you take a template or a style, the prompt goes to the model word for word, without translation or "improvements". This matters for lettering on posters and packaging: translation would rewrite the text in quotes. A free prompt without a template takes the usual path with translation, as everywhere in the studio.

Two more ready-made tools for edits. The "ChatGPT Commands" action runs on the previous engine version: 200 commands for retouching, light, background and posters. The Reve editor is for edits like "remove this object, leave the rest alone": it finds objects in the image itself, you put edits on them, and one instruction is sent. In Telegram the same is available in the @Pixynaibot bot: the images section, the ChatGPT group.

Sign up by email, Google or Telegram, no card required; the starting tokens are enough for the first images.

What changed in ChatGPT Images 2.5 vs Images 2.0

Images in ChatGPT have long been counted in billions per week, and at that volume any small problem becomes a mass one. The 2.5 announcement addresses the four most common ones.

Problem in Images 2.0	What Images 2.5 promises
Editing one object changed everything around it: light, background, face	It edits only what was asked and keeps the rest "even on complex subjects and backgrounds"
After several edits in a row the image "drifted" and lost quality	Each next edit builds on the previous one, earlier changes hold
A person from a reference photo came out similar "in general" but not recognizable	Subjects from photos carry over into new scenes and styles more recognizably, characteristic features are kept more often
Complex layouts, infographics, transparent background came with errors	Support for complex layouts and transparent background is claimed, numbers and labels inside the image are rendered more accurately

About speed: for Flare OpenAI claims up to 50% lower latency compared with Images 2.0. This is OpenAI's measurement on their prompts, there is no independent check. Let us also clear up the naming confusion. In the app the versions are called Images 2.0 and Images 2.5, and in the API the same models are called GPT Image 2 and GPT Image 2.5.

What has not changed: the model still generates a raster image, not layers. It writes text, and writes it well, but if you need an editable layout with fonts, that is a job for a design editor, not a neural network.

Sketch: a neural network turns a sketch or a draft into an image

Before, the composition idea had to be described in words: "a window on the left, an armchair in the center, a floor lamp on the right". Now you can draw it. In ChatGPT it is a separate Sketch button in the "Images" section: it opens a canvas over the chat.

What is inside the canvas, captured on September 8:

• a pen with adjustable thickness, an eraser, text, shapes;
• a palette of 13 colors plus a custom color picker;
• undo and redo;
• an "Attach sketch" button: the drawing goes to the chat as a reference, and you add a text description of the style and details to it.

You do not need to be able to draw. The model reads the sketch as a layout scheme: what stands where, how big it is, which way it faces. Style, light and materials are set by the text. In its own prompting guide OpenAI frames the task like this: keep the layout, proportions and perspective of the drawing, pick realistic materials and light, do not draw anything new. Applied to your scene, it looks like this:

Turn the attached sketch into a photorealistic image. Keep the exact
layout, proportions and perspective of the drawing. A window with daylight on the left, an
armchair in the center, a floor lamp on the right, a rug with a geometric pattern on the floor. Style: an interior photo
for a catalog, warm light, neutral beige tones, no people. Do not add new objects
or text.

In the "ChatGPT Images" editor the canvas works the same way, and the sketch-to-photo style inserts the beginning of this prompt: you only need to describe the scene and the materials. It also works without the built-in canvas: draw the scheme in any editor or on paper, take a photo and attach it as a reference. In the Pixyn studio that is a regular image upload to the prompt.

The prompt formula OpenAI recommends for 2.5

In the prompting guide for GPT Image 2.5 OpenAI has one core scheme: the scene, then the subject, then the details, then the constraints. For complex requests it is convenient to write these four parts as separate labeled paragraphs rather than one long sentence. Three more rules from there that immediately improve the result:

• Write text on the image in quotes and say how many times it should appear: "the slogan exactly once, large, no other text". Spell unusual names letter by letter.
• Give references roles and numbers: "Image 1 is the scene, Image 2 is the dog, put the dog next to the woman, change nothing else".
• For edits, separate "what to change" and "what to keep". The formula "change only X, keep the face, pose, light, background and framing" gives a predictable result where "make it nice" gives something new every time.

Both the templates in ChatGPT and the templates in the Pixyn editor, which we turn to next, are built on this scheme.

Templates: 16 presets that set the prompt structure

The Templates tab in the "Images" section has 16 templates: Poster, Interior design, Logo, Illustration, Headshot, Icon, Product photo, Merch, Infographic, Flyer, Book cover, App mockup, Packaging, Advertisement, Thumbnail, Website mockup. Separately, in the "Popular" tab of ChatGPT itself, there are about forty more style presets: stickers, caricature, anime, a 3D avatar, a studio portrait, a blueprint-style poster, film frames and others.

A template and a preset work differently. Clicking a template in ChatGPT does not open a form. It inserts a ready prompt skeleton into the input field. For Poster it looks roughly like this: "create a poster that conveys the user's idea through a clear hierarchy and readable required text; keep all provided details exactly, do not invent names, dates, places or logos; ask clarifying questions about the message, required content, layout and style". In other words, the template brings discipline: the model first asks what should be on the poster and only then draws it. A preset from "Popular" is simpler: it adds a style.

In the "ChatGPT Images" editor the same 16 templates are made as forms: the poster has three fields (message, required text, style), the logo also has three (name, line of business and character). There are no clarifying questions, the skeleton with constraints is assembled automatically, and an unfilled required field simply does not let you start the generation. That is cheaper than a mistake: without the required text the model would have invented its own.

You can also write such a skeleton by hand, and it carries over to any service with GPT Image. A working prompt for a product card:

Make a marketplace product photo from the attached shot of a mug.
Keep the shape, color and print of the mug exactly as in the photo, do not draw anything on the
product itself. Replace the background with a plain light gray studio one, add a soft shadow below.
Vertical 2:3 frame, the product in the center, taking up 70% of the height. No text or watermarks
on the image.

Note two lines: "keep exactly as in the photo" and "do not draw anything". This is exactly what the improved reference handling in 2.5 is responsible for: without an explicit ban the model often "improves" the product, and on a marketplace that is already an inaccurate card.

For comparison, the prompt "Make a beautiful photo of a mug in a premium style" will give you a mug, but some mug of its own: a different print, a different handle, a different shade. It will do for an article illustration, but a product card needs the first option with explicit bans.

Comments on the image: a precise edit without retelling the whole scene

The third feature sounds modest, but in practice it saves the most time. On a finished image in ChatGPT you put a comment on a specific area: "make this green", "remove this person", "replace the text on the sign with 'Open'". The model changes only the marked area. Before, you had to describe in words where exactly the edit was and hope it would not touch the neighbors.

For the API, the equivalent of such an edit is built from two parts: an instruction describing the place ("in the top left corner", "on the sign above the door") and, if you need a hard boundary, a mask. In the Images API the mask is a PNG of the same size as the source: transparent pixels mark the edit zone. This is how both GPT Image 2 and 2.5 work. In Pixyn the mask is attached in the studio, and object edits without drawing a mask are more convenient in the Reve editor: it finds the objects itself and collects the edits into one instruction.

An edit prompt without a mask that keeps the rest of the scene:

In the attached photo, change only the color of the sofa from gray to dark green velvet.
Keep the light, shadows, the position of the cushions, the walls, the floor and all other objects unchanged.
Do not change the framing or the proportions.

The "keep unchanged" tail seems redundant, but it is exactly what holds the composition. We use the same tail in the Reve editor, where edits on objects are collected into one instruction, and without it the scene noticeably "drifts".

Multi-step edits: why the fifth edit is now no worse than the first

In a dialogue with the model, edits accumulate: first the background, then the light, then the clothes, then the text. With Images 2.0, by the third or fourth edit the quality dropped noticeably and early changes were lost. For 2.5 OpenAI separately claims stable results over a long dialogue: each edit builds on the previous one without degradation.

How to make good use of it:

1. Do not try to describe everything in one prompt. First the composition and the main object, then the light, then the details.
2. Phrase each edit as a difference ("do X"), not the whole scene again.
3. Save a version after each step. If the fifth edit went the wrong way, roll back to the fourth instead of starting over.

The version timeline in the "ChatGPT Images" editor is made exactly for this: each edit comes from the active version, any previous one opens with a click and becomes the source for the next step. In the studio the same technique works through the history: you can always return to a good generation and take it as a reference.

API: Flare vs Sunburst

For developers and services there are two models with the same per-token rates and a different character.

	GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
What it is according to OpenAI	A small model optimized for speed, quality at the level of GPT Image 2	The base model optimized for quality, above GPT Image 2
What for	Everyday generation, social media content, product scenes, prototypes, large volumes	Premium visuals, advertising campaigns, product shoots with repeated edits
Speed (according to OpenAI)	Noticeably faster than GPT Image 2	Slower than Flare, generation takes longer
Edit precision	Improved relative to GPT Image 2	Improved more: OpenAI advises starting with it when quality is critical
Endpoints	generations, edits	generations, edits
Quality	low, medium, high, xhigh, max, auto	the same
Sizes	standard plus custom up to 3840 per side	the same
In Pixyn	"Fast" mode, from 4 tokens	"Best quality" mode, from 4 tokens

What is new in the parameters if you have already worked with GPT Image 2:

• Quality xhigh and max. Two levels above the previous high. Older models do not have them.
• Custom size. Width and height are multiples of 16, the ratio is from 1:3 to 3:1, no side is larger than 3840 pixels, and the total pixel count is from 655,360 to 8,294,400. OpenAI marks everything above 2560×1440 as experimental. The exact rules are in the image generation guide.
• Streaming. The partial_images parameter returns up to three intermediate frames, but each one costs an extra 100 output tokens.
• Mask. The same format and size as the source, with an alpha channel. The edit is applied to the first image.

OpenAI does not publish a separate "price per image" table, only a calculator: the cost depends on size and quality because it is counted in output tokens. The practical takeaway: low for drafts and idea iteration, medium for working images, high and above only for the final. In Pixyn these levels are hidden behind the resolution: 1K, 2K and 4K; the price grows with the resolution.

Which size and quality to choose

Task	Size	Quality
Idea iteration, 5 composition options	1024×1024	low, the closest in Pixyn is 1K
A product card, a blog illustration, a social post	1024×1536 or 1536×1024	medium, in Pixyn 1K
A print poster, a cover, a website banner	2048×2048 or the 3:2 and 2:3 pairs	high, in Pixyn 2K
The final campaign visual, a large format	up to 3840 per side	xhigh or max, only on 2.5; in Pixyn 4K (beta)

Two rules that will save tokens. First: start with 1K, and run the same prompt in 2K only when the composition works. Second: the size for the task, not "bigger". The price is counted in output tokens, their number grows with the size, so a 2K poster costs more than a 1024 square at the same quality.

File format: where photo edits break most often

The model accepts JPEG, PNG and WebP as input. Photos from messengers and iPhones often arrive in AVIF and HEIC, and the provider rejects such a file before generation, while the person sees an unclear error. In our statistics over 90 days GPT Image had ten failed runs, and in nine of them the cause was the input format, not the prompt.

The "ChatGPT Images" editor converts such files to PNG itself, before sending. If the browser could not open a HEIC (this happens outside Safari), the editor asks you to save the file as JPEG or PNG instead of charging tokens for a run that is bound to fail. The size limit of one file: 25 MB.

What to consider before generating an image in ChatGPT

• The result is always raster. Fonts, layers and vector logos have to be moved to an editor by hand. A logo with clean vectors is easier to make in Recraft: it outputs SVG.
• Print. A maximum of 3840 pixels per side, and OpenAI marks everything above 2560×1440 as experimental. For a wide format, generate 2K and add upscaling.
• Labeling. OpenAI adds C2PA metadata and an invisible watermark to every image, as it states in the image generation guide.
• Stream and final. For hundreds of images take Flare, and keep Sunburst for the final edits.
• Neighbors in the catalog. For quick photo edits with an emphasis on naturalness the catalog has Nano Banana Pro, and for lettering on logos and packaging Ideogram.

Frequently asked questions about GPT Image 2.5 and images in ChatGPT

How is GPT Image 2.5 different from ChatGPT Images 2.5? It is one model with two names. In the ChatGPT app itself the feature is called ChatGPT Images 2.5. In the API for developers and services the same model comes out under the names GPT Image 2.5 Flare and Sunburst.

How do I open GPT Image 2.5 in Pixyn? Two ways. In the studio you choose the ChatGPT group and version 2.5 Flare or Sunburst: a regular generation chat with history. In the "ChatGPT Images" action the same models come with a version timeline, references, a sketch canvas, templates and styles.

How is the "ChatGPT Images" action different from the studio? The studio works as a chat: a prompt, settings, the result in the history. The action is built around one image: edits from the active version, a sketch as the source, templates with fields and styles, a mask in a separate zone. For one image that you refine with edits, the action is more convenient; for a series of different images and switching between models, the studio is.

Is Images 2.5 available on the free ChatGPT plan? According to OpenAI's announcement, the update is rolling out to all ChatGPT users, including the free plan, as well as ChatGPT Work and Codex.

Can I use GPT Image for free? In ChatGPT you can, but with the limits of the free plan: OpenAI does not say in the announcement exactly how many images per day. In Pixyn you get starting tokens at sign-up, enough for the first GPT Image images, and after that images are paid with tokens according to the plan.

Can I edit my own photo and not a generated image? Yes. In ChatGPT you upload a photo and describe the edit or put a comment on an area. Through the API the edits endpoint works: you send your image, an instruction and, if you need a hard boundary, a mask. In Pixyn you attach the photo to the prompt in the studio or upload it as the source in the action; the mask is available in the studio.

How do I make an image from a draft or a sketch? The neural network finishes the draft into a ready image. In ChatGPT: the Sketch button in the "Images" section, you draw, attach and add a style description. In the "ChatGPT Images" action: the "Sketch" button, you draw on the canvas, press "Attach sketch" and describe the scene. In any other service: draw the scheme anywhere, attach it as a reference and write "use the sketch as a composition scheme".

Flare or Sunburst, which to choose? Flare by default: it is faster and fits almost everything. Take Sunburst for work where the precision of every edit in a long chain matters: advertising visuals, product shoots for the final.

Does the model keep the faces from my photo? Better than Images 2.0: OpenAI separately claims recognizability of subjects from references. Nobody promises a hundred percent match; for cards with people, check every generation.

What happens to the older GPT Image 1.5 and 2 in the API? They remain available. The difference is in speed, in the xhigh and max quality levels and in edit precision: all of this is only in 2.5.

---

The techniques from the article work today: the ChatGPT group in the Pixyn studio for regular generations, the "ChatGPT Images" editor for edits with versions, sketches and templates. Ready-made commands for retouching and posters are collected in the "ChatGPT Commands" action, and their breakdown with examples is in the article ChatGPT commands for photo editing. The capabilities of the whole group are described on the ChatGPT for images page.

The facts were checked against the ChatGPT Images 2.5 announcement and the GPT Image 2.5 prompting guide in the OpenAI documentation.