The most annoying thing about AI image editing has always been the collateral damage: you ask for a different background and the AI quietly redesigns your product, your model’s face, and the lighting too. ChatGPT Images 2.5, launched September 8, 2026, is OpenAI’s fix for exactly that — an image model built around precise, surgical edits that change only what you ask for while everything else holds still.
It’s faster, better with reference photos, and adds genuinely useful creation tools. Here’s what changed and how to put it to work.
Precise Editing: Change One Thing, Keep Everything Else
The headline upgrade: Images 2.5 is designed to edit only the requested element while preserving the surrounding subjects, composition, and visual treatment. Ask it to change a product photo’s background and it keeps the product, the angle, and the lighting intact instead of regenerating the whole scene.
OpenAI says the model is also better at multi-turn consistency — earlier edits are more likely to survive subsequent instructions, and image quality degrades less as you refine the same asset over multiple rounds. For anyone iterating on thumbnails, product shots, or brand assets, that’s the difference between a usable workflow and starting over every third edit.
Independent early testing has backed up the direction: stronger preservation of individual likenesses and more targeted edits, though results still vary with image and instruction complexity.
Reference Photos That Actually Stay Recognizable
The second big upgrade is reference-photo fidelity. Upload a picture of a person, pet, or product, ask ChatGPT to place it in a new setting or style, and the model now preserves distinctive features far more reliably. Fewer cases of a person coming back looking noticeably different from the original photo.
This matters enormously for creators and sellers: consistent product shots for an Etsy listing, a recognizable mascot across a content series, a personal brand photo restyled without losing your face. Our guide to using AI for Etsy digital products covers workflows where this kind of consistency is the whole ballgame.
New Creative Tools: Sketch, Templates, Comments
Images 2.5 isn’t just a better model — it ships with new ways to direct it:
- Sketch lets you draw a rough visual reference directly inside ChatGPT instead of describing everything in words. On mobile, type @, select Sketch, draw your composition, and add written instructions. Perfect for when positioning or layout is easier to show than explain.
- Templates give starting points for common formats — posters, flyers, product photos — which you then customize with your own content and images.
- Image comments let you point at a specific spot on the picture and describe the change right there, instead of describing locations in text (“the third window from the left”).
Faster: Up to 50% Lower Latency
OpenAI says image generation latency is down by up to 50% compared to Images 2.0. That sounds like a spec-sheet bullet, but for iterative creative work it’s transformative — faster turnaround means more experiments per session, which means better final images. OpenAI reports users now generate more than 3 billion images per week across ChatGPT Images and its API models.
The model is available on all ChatGPT plans (plus ChatGPT Work and Codex), on desktop, mobile, and web. OpenAI didn’t announce any change to generation limits.
For Developers: Flare vs. Sunburst
The same day, OpenAI released two API models:
- GPT-Image-2.5 Flare — the default: fast, high-quality everyday generation
- GPT-Image-2.5 Sunburst — the precision option for workflows where editing control matters most, like multi-turn studio revisions
Both are listed at GPT Image 2’s token rates ($8 per million image input tokens, $30 per million image output tokens), but tokens-per-image differ, so measure your real usage before switching. One catch: neither supports the Batch API — if you run bulk jobs through batch, stay on GPT Image 2 for those.
Practical Uses for Creators
Where this actually pays off:
- Thumbnails — iterate on one winning layout across dozens of videos without the design drifting
- Product photography — restyle backgrounds and settings while the product stays pixel-consistent
- Brand assets — keep mascots, logos-in-scene, and characters recognizable across a whole campaign
- Content series — generate episode art where the setting changes but the subject doesn’t
The through-line: 2.5 rewards creators who think in iterations rather than one-shot generations. Build a base image you like, then refine it — the model now keeps up.
Images 2.5 vs. the Alternatives
How does it stack up for a creator choosing an image tool in late 2026?
- Vs. Midjourney: Midjourney still leads on raw artistic style and community-driven aesthetics, but Images 2.5 wins on precise editing — telling Midjourney to “change just the cup” remains a gamble, while 2.5 is built for it.
- Vs. Adobe Firefly: Firefly’s advantage is commercial safety (trained on licensed Adobe Stock content). Images 2.5’s advantage is editing precision and the conversational workflow inside ChatGPT. For client work where licensing matters, many creators generate in ChatGPT and do final commercial passes in Adobe’s ecosystem.
- Vs. Gemini image generation: Google’s models are strong and deeply integrated with the Google workspace, but OpenAI’s multi-turn editing consistency and the Sketch feature give 2.5 an edge for iterative design work.
The honest answer: most serious creators end up using two tools — one for generation, one for precision editing. Images 2.5 is making a strong case to be the second one.
Getting the Best Results: Prompting Tips
Precision editing rewards precise instructions. A few habits that work well with 2.5:
- Name the target explicitly — “change the coffee cup from white to red” beats “make the cup red,” which beats “fix the cup”
- State what must NOT change — “keep the person’s face, the lighting, and the background exactly as they are” reinforces the model’s preservation behavior
- Use image comments for location — pointing at the exact spot beats describing it in words, especially in busy scenes
- Iterate in small steps — one change per turn preserves quality better than five changes at once, and 2.5’s multi-turn consistency is built for this rhythm
- Lock a reference early — generate your base subject image first, confirm the likeness, then build variations from it rather than regenerating from scratch each time
Official source: OpenAI
Frequently Asked Questions
What is ChatGPT Images 2.5?
OpenAI’s image generation and editing model launched September 8, 2026, focused on precise editing, reference-photo fidelity, multi-turn consistency, and up to 50% faster generation than Images 2.0.
How is the editing different from previous versions?
Images 2.5 changes only the element you ask about while keeping the rest of the image intact — previous models tended to regenerate the whole scene, altering things you wanted left alone.
What is the Sketch feature?
Sketch lets you draw a rough composition directly in ChatGPT as a visual guide for the image you want, instead of describing layout entirely in words. On mobile it’s accessed by typing @ and selecting Sketch.
What’s the difference between Flare and Sunburst?
Both are API models released alongside Images 2.5. Flare is the fast default for everyday generation; Sunburst is tuned for maximum editing precision in demanding multi-turn workflows. Token rates match GPT Image 2.
Is Images 2.5 available on the free plan?
It’s rolling out across all ChatGPT plans, plus ChatGPT Work and Codex, on desktop, mobile, and web. OpenAI hasn’t announced changes to generation limits.
Can it keep a person’s face consistent across edits?
It’s significantly better at preserving subjects from reference photos through successive edits, according to both OpenAI and early independent testing — though complex images and instructions can still vary.