What is GPT Image 1.5, the model behind the new ChatGPT Images?

Yesterday, Tuesday, December 16, 2025, OpenAI released a new version of ChatGPT Images, powered by a new image model. In the API, the developer interface, the model is called GPT Image 1.5. It is rolling out to every ChatGPT user, free and paid, and to developers worldwide. OpenAI's pitch is simple: you should get the picture you had in mind, and when you ask for a change, only that change should happen.

The update has two parts: a better model, and a new place in ChatGPT to use it. The model works with any chat model, so you don't need to pick anything special to try it.

What changed

OpenAI describes these main improvements over its first image model, which it released earlier this year:

Area	What OpenAI says improved
Editing	Changes only what you ask, while lighting, composition and people's appearance stay consistent, even across a series of edits
Instructions	Follows detailed requests more reliably, including where each element belongs in a complex scene
Text	Renders denser and smaller text, useful for posters, menus and infographics
Speed	Up to 4x faster generation; a new request can begin before earlier ones finish

There is also a new Images space in the ChatGPT sidebar, on chatgpt.com and in the mobile app. It offers dozens of preset filters and trending prompt ideas, so you can transform a photo without writing a prompt at all. OpenAI says it will refresh these presets regularly. Business and Enterprise accounts will get this new space later. The previous image version remains available as a custom GPT for anyone who prefers it.

For developers, GPT Image 1.5 brings the same gains to the API, and image inputs and outputs cost 20% less than with GPT Image 1. OpenAI highlights better preservation of logos and brand visuals across edits, which matters for marketing teams and online shops that need many product shots from a single photo. Companies such as Wix, Canva, Figma and Shutterstock are named as early users.

How AI image editing works, and why it's hard

An image model generates a picture from a description, building it up so that it matches the words you gave it. Editing is harder. When you upload a photo and ask for one change, the model has to produce a new image that differs in exactly that one way and matches the original everywhere else.

Many image tools fail here. Ask them to "make the lighting colder" and they quietly redraw the whole scene: a face shifts slightly, the background changes, a logo gets garbled. TechCrunch notes that this lack of consistency has made most image tools poor at iteration, the back-and-forth of refining a picture in small steps.

OpenAI says GPT Image 1.5 handles several kinds of edits while protecting the rest of the picture: adding something, removing something, combining elements from several images, blending styles, and moving an element from one picture into another. OpenAI has not published technical details of how the model is built, so the best way to judge these claims is to test them.

OpenAI is also open about limits. Rerunning examples from its first launch, it found clear progress but imperfect results: some artistic styles, scenes with many faces, and text in languages that don't use the Latin alphabet still need work. In one of its own examples, a poster of deep-sea creatures was about 70% scientifically correct.

A worked example: editing in steps

Here is how a small business owner might refine a product photo, one change at a time, in the same conversation:

Prompt

1. [Upload a photo of a ceramic mug on a kitchen counter]
   Remove the clutter behind the mug. Keep the mug, its shadow and the
   lighting exactly as they are.
2. Change the counter to light oak wood. Do not change the mug.
3. Add the words "Handmade in Quebec" in small white lettering in the
   lower right corner.
4. Make a square version for social media, with the mug centered.

Each step names what should change and what must stay the same. If a step goes wrong, say exactly what drifted ("the handle changed shape; restore the original handle") instead of starting over.

Try it: Upload a photo of a room in your home and ask ChatGPT to change only one thing, such as the color of one wall. Compare the result with the original side by side, and list anything else that changed without being asked.

Why it matters

The timing is not an accident. TechCrunch reports that OpenAI had planned a new image model for early January and moved it forward, in a race with Google, whose Gemini 3 and Nano Banana Pro image model have topped public leaderboards. Last week OpenAI released GPT-5.2 with the same goal of regaining the lead.

There is also a shift in what image tools are for. With reliable editing, an image model becomes closer to a photo studio than a slot machine, useful for product pictures, ads and presentations, not just novelty art. Fidji Simo, OpenAI's CEO of Applications, described the new sidebar space as working more like a creative studio, and she said ChatGPT will also show more visuals in ordinary answers when a picture explains something better than words.

How do you get precise edits from the new ChatGPT Images?

Be explicit about two things in every request: what to change and what to keep. "Change the jacket to red; keep the face, pose, background and lighting the same" works far better than "make it red," because it tells the model which parts of the picture it is allowed to touch.

Work in small steps rather than one long list of changes. Make one edit, check it, then make the next. If you need text in the image, put the exact words in quotation marks and say where they go, then check every letter, since small text is improved but not flawless. For a multi-image combination, say which image each element comes from.

Finally, keep your original file. If a series of edits drifts, it is often quicker to go back to the source photo and repeat the steps with clearer instructions than to fix a picture that has already changed in ways you didn't intend.