You can now try ChatGPT Images 2.5, the new offering in image generation from the most popular AI

In April of this year, OpenAI launched ChatGPT Images 2.0 with the clear intention of stealing the spotlight from Gemini and its Nano Banana 2, and now the company has unveiled its next version, ChatGPT Images 2.5. And no, this is not just a cosmetic name change; there are meaningful underlying updates worth unpacking, especially if you work with visual content, marketing, or design and need to know whether it is worth updating your workflow (spoiler: probably yes).

What is ChatGPT Images 2.5?

It is OpenAI’s new image generation and editing model, available now in ChatGPT, ChatGPT Work, and Codex, across desktop, mobile, and web. To put its scale into perspective, OpenAI itself says that its tools (between ChatGPT Images and the GPT-Image API models) generate more than 3 billion images every week. It is honestly hard to picture a number like that, but it makes it clear why every improvement, no matter how small it may seem, ends up having a massive impact at scale.

According to the official announcement, the promise of this version can be summed up in four words: more detail, more speed, more accuracy, and better tools for creating and sharing. Let us take a closer look at what is behind each of those.

The real differences between Images 2.0 and Images 2.5

It generates images up to 50% faster

This is probably the change that will be most noticeable day to day. OpenAI says it has reduced generation latency by up to 50% compared with Images 2.0. For someone who uses the tool only occasionally, that may not sound especially important, but if you work by generating variations, testing prompts, or iterating on the same concept, that time reduction translates into less friction and more ideas tested per hour.

It edits only what you ask for, without touching the rest

This is, to me, the most interesting leap. One of the biggest headaches in AI image editing, and anyone who has tried to change only the background in a photo and ended up with a different face knows this, is when the model “recreates” parts it should not touch. Images 2.5 is trained to make surgical edits, changing a product, some text, or a background while keeping the rest of the composition intact, even in complex scenes.

For example, ChatGPT Images 2.5 does not alter a person’s face when you ask it to modify only another element in the image. And if you have ever tried to do that with an earlier version, you know this removes a lot of frustration. 

It maintains consistency in long editing conversations

In Images 2.0, we could already generate up to eight images in a single batch while maintaining a certain level of continuity across characters and objects. Images 2.5 goes a step further: during long conversations, where you keep requesting adjustment after adjustment, the model does a better job of preserving previous changes instead of gradually losing quality or “forgetting” earlier edits. If you create content series for social media or campaigns with multiple assets derived from the same base, this will save you a fair amount of time and back-and-forth with the chat.

Greater fidelity when using reference photos

Another improvement that OpenAI emphasizes is what happens when you upload a real photo of a person, a pet, or a place to use as a base. According to the company, this model does a better job of preserving the subject’s distinctive features, even when you place them into completely different styles, settings, or compositions.

Understanding complex instructions and transparent backgrounds

Images 2.5 also does a better job of understanding detailed visual instructions: layouts with specific hierarchy, content with more accurate real-world data, and now support for transparent backgrounds in a more consistent way, which is essential for anyone who needs ready-to-use assets for templates, presentations, or branded materials without having to manually cut them out afterward.

Sketch, templates, and sharable prompts: the new tools arriving with this version

Beyond the model itself, OpenAI added three new features directly into the ChatGPT interface that appear to compete head-on with Google Pics:

• Sketch: you can draw a sketch directly in ChatGPT and use it as a visual guide for the AI to turn into a finished image. It is activated by typing “@Sketch.” Sometimes it is easier to draw an idea badly than to describe it well with words.
• Templates: instead of starting from a blank page, you can now choose popular formats such as “Poster” or “Merch” and simply fill in the details you want to customize.
• Image comments and sharable prompts: you can point directly to the image to show what you want changed (instead of describing everything in text) and share the exact prompt so someone else can use it with their own photos and data.

GPT-Image-2.5 Flare and Sunburst: what changes for developers

For those building products on top of the OpenAI API, the update introduces two new models, each designed for a different use case:

• GPT-Image-2.5 Flare: this is the default option for most applications. It delivers higher-quality images than GPT-Image-2 with up to 50% less latency, making it ideal for creator content, social media, visual search, or high-volume generation.
• GPT-Image-2.5 Sunburst: designed for workflows where fine-grained control in each edit matters more than speed, such as polished production campaigns or premium product imagery.

Companies such as Adobe, Runway, Manus, and Higgsfield AI have already integrated these models into their own platforms. Adobe, for example, added them to Firefly, highlighting the consistent resolution maintained even after several refinements on the same image, according to its product team in OpenAI’s official announcement.

Watermarking, C2PA metadata, and the regulatory context

All images generated with Images 2.5 include C2PA metadata and an invisible watermark that helps identify AI-created content. This comes in response to the fact that, as of August 2, 2026, the requirement to label AI-generated content under the European Union’s AI Act came into force.

When can I use it and how much does it cost?

The good news is that there is no fine print this time: Images 2.5 is already available to all ChatGPT users, across all plans, including ChatGPT Work and Codex, on desktop, mobile, and web. It is not a Plus- or Pro-exclusive feature as happened with some advanced reasoning capabilities in Images 2.0.

For API users, GPT-Image-2.5 Flare and Sunburst operate on a pay-as-you-go model, with pricing varying based on the resolution and complexity of each generation. You can check the latest details on OpenAI’s official pricing page.

Is it worth making the jump from 2.0 to 2.5?

If you use ChatGPT Images casually, to generate the occasional standalone image, you probably will not notice a dramatic change beyond the speed. But if your daily work involves iterating on the same asset, adjusting a product in a campaign, maintaining consistency across multiple images for the same brand, or simply avoiding the AI “ruining” someone’s face when all you wanted was to change their T-shirt, this update does address those frictions.

What is clear is that the race between OpenAI and Google in AI image generation is not going to slow down anytime soon. Since Nano Banana arrived in August 2025 and cemented Google’s lead in image editing, OpenAI has been responding blow for blow. Images 2.5 is, so far, its strongest response.

Image: GPT