Your product image is almost ready. You ask for a different background, and suddenly the bottle and label have changed too. If you create images with AI, that cycle of revisions probably feels familiar.
OpenAI released ChatGPT Images 2.5 on September 8, 2026. Its focus is on making those revisions more useful, with better subject preservation, more accurate edits, and shorter generation times. Read the official announcement.
ChatGPT Images 2.5 support is coming soon to HeyMarmot. Get your reference images and creative briefs ready. We will share the available versions, credit pricing, and release timing in our launch announcement.
What Is New in Images 2.5?
According to OpenAI, the key improvements include:
- Better image fidelity: more natural lighting and textures, with improved preservation of subjects from reference photos.
- More precise edits: better at changing the requested element while retaining surrounding details.
- More consistent revisions: improved continuity across multiple edits, with less image quality degradation.
- Shorter waits: up to 50% lower generation latency compared with Images 2.0. This is OpenAI's reported comparison; actual timing depends on the task.
ChatGPT also adds Sketch references, templates, image comments for editing, and prompt sharing. These are ChatGPT product features; HeyMarmot's supported features will be detailed at launch. Source: OpenAI's release announcement.
For creators, the useful question is whether a few deliberate revisions can bring an image closer to a finished asset. You might settle the composition first, adjust the lighting next, and update the copy last. Each step has a clear purpose and a result you can inspect.
Flare vs. Sunburst: What Is the Difference?
The API release includes two versions. Here is how their positioning maps to creative tasks:
| Version | Official positioning | Tasks to consider |
|---|---|---|
| GPT Image 2.5 Flare | Fast, high-quality everyday image generation | Social visuals, early concepts, and quick creative exploration |
| GPT Image 2.5 Sunburst | Image generation and editing focused on precision | Detailed product images, campaign assets, and final revisions |
Flare accepts text and image inputs and produces images. Sunburst also accepts text and images, with an emphasis on workflows where editing precision matters most. Both offer low, medium, high, xhigh, max, and auto quality settings. Flare documentation, Sunburst documentation.
Both are available through the Image API and the Responses API image generation tool. Their token rates match GPT Image 2, but the cost of an individual image still depends on actual token usage. That does not establish a fixed price per image. API changelog, official model pricing notes.
Three Creative Workflows to Try
The following ideas and prompts are our own starting points for exploring the release. They are not results from HeyMarmot testing.
1. Give the Same Product a New Setting
Product imagery often needs to do two things at once: make the scene feel fresh and keep the item accurate. Start with a clear product photo and specify the packaging, text, and shape that must stay intact.
Using my uploaded product photo as a reference, create a skincare promotional image. Preserve the bottle shape, cap color, brand logo, and label text. Replace the background with a pale beige stone countertop, lit by soft natural light from the upper left. Place the product on the right side of the frame and leave room for a headline on the left. Do not add other products or new text.
Review the product itself before deciding whether to adjust the background again. This makes it easier to spot which revision introduced an unwanted change.
2. Make One Specific Edit to a Poster
When a poster is nearly finished, a useful instruction can be short: identify the location, describe the replacement, and specify what should remain.
Change only the event date in the upper-right corner of the poster to “September 20.” Preserve the headline, person, background, type style, and position of every element. Use the original date's font size and color.
Check dates, prices, and brand names character by character before publishing. Better editing performance still leaves a role for a final proofread.
3. Refine a Keyframe Before Making a Video
A still image can help establish a character, location, and camera angle before video production begins. Save each approved revision so you can compare versions or return to an earlier direction.
Use this character image as a reference for a static storyboard frame opening a short film. The character stands outside a convenience store at night after rain. Warm interior light falls across their face, and the wet pavement reflects the store sign. Preserve the hairstyle, jacket, and backpack from the reference. Use a medium shot with the character on the left and space for the street on the right.
Once the composition works, change one factor at a time. For example: “Reduce the reflections on the pavement and keep everything else unchanged.” A sequence like this also gives you a concrete way to compare how different models handle the same brief.
Coming Soon to HeyMarmot
We are bringing ChatGPT Images 2.5 to HeyMarmot soon. If you have product visuals, social content, or video storyboards in mind, you can prepare three things now:
- Clear reference images: a product photo, character reference, or original image you want to edit.
- A defined use: the publishing channel, aspect ratio, and space needed for copy.
- Details to preserve: brand text, clothing, product structure, or an approved composition.
At launch, try it with a real creative brief. A specific task makes it much easier to judge whether the results fit your work than a broad request for a beautiful image.
Watch the HeyMarmot workspace for the launch. In the meantime, explore our GPT Image 2 prompting guide to prepare prompts for your next project.