ChatGPT Images 2.5, twice the speed and precise element-level editing

Three Billion Images a Week

The figure OpenAI opens its announcement with says a lot on its own. Every week, people create more than three billion images through ChatGPT Images and the GPT-Image models in the API. This isn't an experiment anymore, it's a mass-market tool.

On September 8, 2026, the company released ChatGPT Images 2.5. There are four main improvements, sharper details, more precise editing, faster generation, and new tools built right into ChatGPT. The model produces more natural lighting and richer textures, does a better job preserving the people and objects in your reference photos, and follows editing instructions more reliably across a long conversation.

The most noticeable change day to day is speed. Generation latency dropped by up to 50 percent compared to the previous Images 2.0, which makes cycling through variations considerably faster.

How well the model preserves people from your photos

The most meaningful images usually involve real people, places, and memories. That's exactly why OpenAI put focused work into how the model handles reference photos.

Images 2.5 does a better job carrying familiar people and objects into new settings, different visual styles, or altered compositions. Subjects look more recognizable in the final image, lighting and textures feel more natural, and distinctive features are more likely to carry through.

For anyone generating images at volume through the API, this means one thing. A series of variations now stays anchored to the original source, so the pictures don't drift apart from frame to frame.

Precision editing, change one thing and leave the rest alone

This is arguably the most practically useful part. Previously, asking to fix one detail often meant half the picture changed. Images 2.5 is noticeably better at editing only what you asked for while keeping the rest of the details the same, even with complex subjects and backgrounds.

For anyone working on advertising or product listings, the benefit is obvious. You can update a single element, like the product itself, the background, or a piece of copy, while preserving the subject, composition, and brand treatment around it.

There's a related new feature in the interface too. You can now place comments directly on an image, pointing to exactly which spot needs reworking. That's easier than describing in words where on the picture the relevant piece sits.

Why quality no longer degrades after the tenth edit

Anyone who has spent a long session refining an image knows the problem. Every new edit slowly degrades the picture, earlier changes get lost, and at some point starting over is easier.

Images 2.5 handles this better. During longer conversations the model follows specific editing instructions more reliably across multiple turns. Earlier changes are more likely to stay in place, and each new edit builds on the work already done without degrading quality over time.

In production workflows this matters a lot, since you often need a targeted change without rebuilding the whole asset. Accumulated edits used to be the weak spot of every image generator.

Complex briefs and transparent backgrounds

Images 2.5 is better at understanding complex visual instructions and turning them into coherent results. Images containing real-world information have more accurate content, and the model handles more complex layouts including transparent backgrounds.

Style handling improved too, so images land closer to what you envisioned artistically.

Here's where that shows up most:

• A series of advertising assets in one consistent brand style

• UI concepts where a given hierarchy of elements must be preserved

• Presentation visuals that need to fit a defined structure

The key point is that the model is more likely to retain the requested visual direction, composition, and individual details instead of drifting as instructions get more specific. Detailed briefs used to be exactly where generators lost the thread.

How hand-drawing works right inside ChatGPT

Sometimes the clearest way to explain an idea is to draw it. That's the whole premise behind the new Sketch feature.

You draw directly in ChatGPT and use the sketch as a visual guide for the final image. You can quickly rough out a room layout, the contour of an outfit you're designing, or just a funny doodle, and ChatGPT turns that rough art into a complete image. Alongside the sketch you add a description of the style and any other details to make sure it reflects your vision.

You don't need to be a professional artist, it's simply another way to get the final image closer to what you have in mind. To try it, just type "@Sketch" in ChatGPT.

Templates and sharing your prompt

Sometimes the idea is clear but getting started is hard. Templates for the most popular creative formats exist for exactly those moments.

Instead of a blank canvas, you pick a ready-made template like "Poster" or "Merch" and add your own details. That might be the information to convey, design elements, or a style, so the result comes out yours.

The second nice touch involves sharing results. When you share an image now, you can include the prompt that got you there. Someone else can pick up the same idea, drop in their own photos and details, and end up with a version of their own.

Two new models for developers

Alongside the ChatGPT update, the company released two models in the API. They differ in how they balance speed against precision.

GPT-Image-2.5 Flare brings the same improvements in quality, editing, and speed, and is the default choice for most applications. It delivers higher-quality images than GPT-Image-2 at 50 percent lower latency. It suits creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation.

GPT-Image-2.5 Sunburst is built for premium visual workflows that need tighter control across edits. Its territory is production-ready campaign creative and polished product imagery, though generation takes longer.

Choosing between them comes down to a simple question. If speed and volume matter, take Flare, if maximum precision on every edit matters and you can wait, take Sunburst.

Safety and watermarking

ChatGPT Images 2.5 builds on existing safeguards, with checks applied to both the prompt and the resulting image to help prevent harmful outputs.

The company continues to use C2PA metadata and invisible watermarking to help identify images made with its tools. Invisible means nothing is visible in the picture to the naked eye, but a dedicated check can detect the mark.

By 2026 this has become standard practice across all major developers, and we've covered similar approaches from other companies before, so the move is entirely expected.

Where to try it and what it costs

Images 2.5 is rolling out today to ChatGPT, ChatGPT Work, and Codex users across all tiers, on desktop, mobile, and web.

The GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare models are available in the API, with pricing published separately on the company's pricing page.

Speed Versus Precision

ChatGPT Images 2.5 addresses the two biggest pain points of every image generator at once. Slow generation that makes cycling through variations tedious, and image degradation across successive edits, where starting over is easier by the tenth iteration. On top of that, hand-drawing through Sketch gives you a way to explain an idea when words fall short.

That said, OpenAI isn't alone in this space. Nano Banana, Midjourney, Flux, and others each have their own strengths, whether in artistic aesthetics, text inside the frame, or cost per image. There's no universal leader here.

There's only one way to find out which model handles your specific task best, and that's running the same request through several generators and comparing the results with your own eyes.

That's exactly what unitool.ai is for. One subscription unlocks access to every current image and video model at once, no foreign cards and no separate sign-up on each platform. Upload your photo, give the same prompt to several models, and pick the one that understood you best.