Portrait retouching
Selfies, ID photos, virtual makeup. Flare makes iteration fast and cheap, and character consistency keeps the person in round five looking like round one.
OpenAI's newest image model, shipped September 2026 as the step up from GPT Image 2
ChatGPT Images 2.5 AI Image Agent
Type one sentence — your first image is on the house, no sign-up, about half a minute. Then 100 free credits when you sign up, and more every day you check in. No card, no ChatGPT subscription.
Before
AfterThe models behind the agent
ChatGPT Images 2.5GPT Image 2Nano Banana ProSeedream 5 ProFLUX KontextVeo 3.1Kling 2.5Seedance 2
What it is
What is the Images 2.5 AI Image Agent?
ChatGPT Images 2.5 — also written GPT Images 2.5 or simply Images 2.5 — is OpenAI's newest image model, released on September 8, 2026. Its headline abilities: text that finally renders correctly inside images, true transparent backgrounds straight from a prompt, characters that stay consistent across successive edits, and frame-level control that keeps a subject coherent across every cell of a sprite sheet. Output runs at 1K, 2K or 4K, in ratios from ultra-tall 1:8 through square to ultra-wide 8:1.
The API ships two variants. Flare is the everyday engine: higher quality than GPT Image 2 at up to 50% lower latency, priced for fast iteration. Sunburst is the precision engine: the best text and detail rendering with tighter control across successive edits, built for production creative. Images 2.5 — this site — puts both inside an agent you can talk to: you write one sentence, it checks the prompt library, searches the web for reference when the subject demands it, writes the detailed English prompt, picks the right variant, and keeps refining in the same chat.
How it works
How to use the AI image agent
Three steps from idea to finished image. The same chat handles text-to-image, image editing and video, so there is nothing else to learn.
Describe the image
Type one sentence in any language — no prompt template needed — or attach photos to edit. The example prompts under the composer are one-click starting points.
Let the agent pick the model
On Auto it matches the engine to the job — Flare for speed, Sunburst for precise text and detail — after checking the prompt library and, when the subject needs it, the web. Then it writes the detailed English prompt and generates.
Refine in the same chat
Ask for changes in plain words: swap the background, fix the lighting, match a reference. Every result stays in the thread, ready for the next edit, a sprite-sheet GIF, or a video clip.
Flow canvas
Every creation unfolds into a flow map
You chat on the left; an infinite canvas on the right draws the whole process in real time. Requests, web research, option forms, prompts, renders and every revision become nodes, wired the way the data actually flowed. Ten rounds in, you can still see exactly how each image came to be.
Steps wire themselves
Searches connect to the prompt, the prompt to the render, and a revision request hangs off the very image it changes — even form answers stay on the graph. Full lineage at a glance.
Hover, then act
Point at any render and fire the next step on the spot — remove background, upscale, make a GIF, generate video, or edit again. The new step grows onto the canvas and the chat stays in sync.
Arrange it your way
Drag nodes anywhere with snap-to-grid and edges that follow. Pan with two fingers, pinch to zoom, double-click empty ground to refit the view.
Full screen, your split
Take the canvas truly full screen, or drag the divider between chat and canvas until each side has exactly the room you want.
Use cases
What people make with Images 2.5
Everyday jobs matched to the engine that does each best — all from the same chat box above.
Covers and posters
Event posters, covers and banners where the headline must read letter-perfect — Sunburst's typography rendering, in any ratio up to ultra-wide 8:1.
Logos and brand marks
True transparent-background PNGs straight from a prompt: logos, sticker sheets and cutout product shots, ready to place on anything.
Style transfer
Repaint a photo as watercolor, anime, pixel art or the viral 1980s portrait look with Flare — seconds per try, so experimenting costs nothing.
Need something else? Ask the agent directly in the chat
Features
One agent, every image job
Images 2.5 is a chat with tools: image generation, editing, GIF assembly and video, each backed by the model that does it best.
Conversational creation
Say it once. The agent consults the site's prompt library, searches the web when the subject needs real reference, expands your sentence into a full English prompt and picks the engine on Auto. No parameter forms, no prompt engineering.
The best model for each request
ChatGPT Images 2.5 Flare and Sunburst lead the catalog, with GPT Image 2, Nano Banana Pro, Seedream and FLUX Kontext behind them. The picker defaults to Auto — the agent matches the engine to the request — and you can pin any model when you care.
ChatGPT Images 2.5
GPT Image 2
Nano Banana Pro
Seedream 5 Pro
FLUX Kontext
Veo 3.1
Editing and composition
Attach one photo to retouch it, or several to combine them: put two people in one picture, try a garment on, place a logo on a product. Images 2.5 keeps faces and characters consistent across successive edits, so round five still looks like round one.
GIFs and video from text or image
Ask for a GIF and the agent uses Images 2.5's frame-level consistency: one sprite-sheet generation, sliced and assembled server-side into a looping GIF — with true transparency. Ask for a clip and a video model renders it, or animates an image from the conversation.
- Veo 3.1 Fast: 8-second clips with sound
- Kling 2.5 and Seedance 2 for motion quality
- PixVerse V6 and Grok Imagine for quick drafts
Why Images 2.5
Why choose this AI image agent
There are many image generators. This one is an agent: it does the prompt writing, the model picking and the retrying, and you keep the conversation.
No prompt engineering
Write one plain sentence in your own language. The agent turns it into a detailed English prompt that models actually follow.
Image and video in one chat
Generate a picture, refine it, then animate it, all in the same thread with the same credits.
True multi-image editing
Combine people, garments, logos and backgrounds from several photos in one instruction.
Failed generations are refunded
A generation that errors out returns its credits automatically. You pay for results.
Clear per-image pricing
Every model shows its credit price before you generate. Cheap drafts, premium finals, your call.
Your images stay yours
Uploads and results live in your private library. Delete them anytime; use results commercially on paid plans.
FAQ
Frequently asked questions
ChatGPT Images 2.5 is OpenAI's newest image generation model, released on September 8, 2026 as the successor to GPT Image 2. You will also see it written as GPT Image 2.5 or Image 2.5 in the singular, as gpt-image-2.5 (the API model id), or shortened to Images 2.5. It produces more natural lighting and richer textures, better preserves the subjects of reference photos, follows editing instructions reliably across multiple turns, and cuts generation time by up to 50%. Images25.ai gives you access to it through a conversational AI image agent.
No. Images25.ai is an independent tool. We access ChatGPT Images 2.5 and the other models in the catalog through licensed third-party APIs, and add the agent experience, the library and the editing workflow on top.
A normal generator hands you a prompt box and settings. An agent reads your request, writes the detailed prompt, picks the model, checks the result and takes your follow-up corrections in the same conversation. It is closer to briefing a designer than operating a tool.
Image models include ChatGPT Images 2.5 in both API variants (Flare and Sunburst), GPT Image 2, Nano Banana Pro and 2, Seedream 5 Pro and 4, FLUX Kontext Pro and Z-Image Turbo. Video models include Veo 3.1 Fast, Kling 2.5, Seedance 2 Fast, PixVerse V6 and Grok Imagine Video. The catalog grows as new models ship.
Yes. Your first image is free without an account, and signing up adds 100 credits — enough for 2 images or 1 video, no credit card required. After that, check in once a day: a 7-day round pays up to 200 more, with a bigger reward every few days. Draft models stretch the credits to many more images, while the premium Images 2.5 engines spend them faster — every model shows its own credit price in the picker.
Each model lists its credit price per image or per clip; premium models cost more than draft models, and a larger image (2K, 4K) costs more than the standard 1K. Credits are only spent on successful generations, and failed ones are refunded automatically.
Yes. Attach a photo in the chat and describe the change: remove the background, fix the lighting, restyle it, or merge it with another picture. You can also draw arrows and circles on the image to point at exactly what should change.
Yes. Ask for a short clip and the agent generates it from your text, or animates an image from the conversation. Clips run a few seconds at 720p to 1080p, depending on the engine, and are priced per clip, with the cost shown before you confirm.
Images and videos you generate on a paid plan can be used commercially. Free-tier results are for personal use. You are responsible for content you create from your own uploads.
The composer offers ratios from ultra-tall 1:8 through square 1:1 to ultra-wide 8:1 (21:9 included), and 1K, 2K or 4K output on models that support it. Or skip the settings and tell the agent "wide banner, 4K" in the chat.
Flare and Sunburst are the two API variants of ChatGPT Images 2.5. Flare is the fast default: higher quality than GPT Image 2 at up to 50% lower latency — ideal for iteration, style transfer and everyday images. Sunburst is the precision variant: the best text and typography rendering, tighter control across successive edits, and the steadiest fine detail — pick it for posters with real copy and production work. On Auto, the agent chooses between them for you.
Yes — that is a signature ChatGPT Images 2.5 trick. The model controls frame-to-frame change, so the agent generates one sprite sheet of consecutive animation frames, slices the grid server-side and assembles a looping GIF, with true transparency when you ask for it. Attach a character and say "a 4x4 pixel-art run cycle of this character as a GIF".
Both. Sunburst renders headlines, labels and typography correctly inside the image — put the exact wording in quotes. And the Images 2.5 models output true alpha-channel transparency from a plain request ("isolated on a transparent background, PNG"), no manual cutout needed.
Every chat has a second view: an infinite canvas that draws the session as a graph — requests, web searches, option forms, prompts, renders and revisions, wired in data-flow order automatically. Drag nodes with snap-to-grid, pinch to zoom, go full screen; hover any render to fire the next step (remove background, upscale, GIF, video, or another edit) and the conversation keeps itself in sync.
Attach any portrait and ask for the eighties look — for example: "Turn this photo into a 1980s studio glamour portrait: big feathered hair, soft-focus glow, a pastel airbrushed backdrop, a denim jacket and subtle film grain. Keep my face recognizable." ChatGPT Images 2.5 keeps the person while restyling everything else, which is exactly what the trend depends on, and on Flare each variation takes seconds — try mall-photo, prom or yearbook versions of the same shot. A ready-made 1980s example sits under Style transfer in the composer.
Yes — developers reach the model through the API under two ids, gpt-image-2.5-flare and gpt-image-2.5-sunburst. Here you need none of that: no API key, no code — the agent calls both variants for you, and you can start on the free credits. If you are building on the API, the chat is also the fastest place to test a prompt against Flare and Sunburst before committing it to code.
Make your first image with ChatGPT Images 2.5
One sentence in the chat is enough. 100 free credits on sign-up, more every day you check in — no credit card required.