Depikt
Back to blog
Images 2.57 min

ChatGPT Images 2.5: What's New and What It Means for Prompting

OpenAI shipped Images 2.5 on September 8, 2026. Here is what actually changed, what is only a ChatGPT feature, and how it changes the way you should write a prompt.

By Depikt Team

The short version

On September 8, 2026, OpenAI announced ChatGPT Images 2.5 and described it as its new state-of-the-art image model, with "sharper details, faster generation, more precise editing, and better tools for creating and sharing." It rolled out to every ChatGPT, ChatGPT Work, and Codex tier the same day.

If you only read one paragraph: the model got noticeably better at three things that used to eat generations, namely keeping the person in your reference photo recognizable, changing only the thing you asked it to change, and not degrading an image across a long editing conversation. Everything else in the launch is either a ChatGPT product feature or an API model name.

This article sorts the announcement into those three buckets, then covers what each change means for the prompt you write. Depikt's Prompt workspace writes and reviews prompts rather than generating the image itself, so the second half is our interpretation, and it is labelled as such.

What changed in the model

These are the claims OpenAI makes about the model itself, paraphrased from the announcement and the system card.

1. Reference fidelity

Images 2.5 is "better at preserving the subjects in your reference photos." OpenAI's examples are all reference-led: a printed baby portrait remixed into a tuxedo, a dog in costume, a photobooth strip turned into a headshot. The claim is that subjects look more recognizable, lighting and textures feel more natural, and distinctive features carry through when you move a subject into a new setting, style, or composition.

2. Precision editing

The model is "better at editing only what you've asked for, while keeping the rest of the details the same," including with complex subjects and busy backgrounds. OpenAI frames the API version of this as updating a single element such as a product, a background, or a piece of copy while the subject, composition, and brand treatment stay put.

3. Multi-turn consistency

In longer conversations, earlier edits are "more likely to stay consistent, and each new edit builds on the work you've already done without degrading image quality over time." This is the quiet one. Anyone who has watched a face slowly become someone else over six edits knows why it matters.

4. Intelligence and layout

Images 2.5 is described as better at understanding complex visual instructions. Images that include real-world information "have more accurate content," the model "can handle more complex layouts including transparent backgrounds," and it is better at reflecting a requested visual style. The system card adds that it "improves infographic accuracy and layout." OpenAI specifically names UI concepts that preserve a hierarchy and presentation visuals that fit a defined structure.

5. Speed

Generation latency is "up to 50%" lower than Images 2.0. In the API, the default model is positioned as delivering higher quality than GPT-Image-2 at half the latency.

What OpenAI did not publish: benchmark numbers for text rendering, any prompt-length guidance, or a side-by-side with other vendors. If you read those claims elsewhere, they are not from the launch.

What changed in ChatGPT (not the model)

Four features shipped alongside the model. They live in the ChatGPT app, and the API does not expose them as such.

  • Sketch. Draw directly in ChatGPT and use the drawing as a visual guide for the final image. You invoke it by typing @Sketch. OpenAI's examples are a room layout, the contour of an outfit, and a doodle.
  • Templates. Pick a format like "Poster" or "Merch" and fill in the information, design elements, and style instead of starting from a blank prompt.
  • Comments on images. Place a comment directly on part of an image to focus an edit there.
  • Shared prompts. When you share an image you can include the prompt that made it, so someone else can rerun it with their own photos and details.

The important thing for prompting: Sketch and comments are new ways to hand the model a spatial reference. They do not replace the words. A sketch tells the model where; the prompt still has to say what, in what style, and what to leave alone.

What changed in the API

Two models, both released the same day and both snapshotted 2026-09-08:

  • gpt-image-2.5-flare is the default. OpenAI describes it as its fastest model for high-quality everyday generation, well suited to creator and social content, product experiences, visual search, prototyping, and high-volume work.
  • gpt-image-2.5-sunburst is for "premium visual workflows that benefit from tighter control across edits," with longer generation times. Think campaign creative and polished product imagery.

The Flare model page lists quality levels of low, medium, high, xhigh, max, and auto, and both image generation and image edit endpoints. If you are not building on the API, none of this changes how you prompt in ChatGPT.

What it means for prompting

From here on this is Depikt's reading of the launch, not OpenAI's guidance.

The payoff for being explicit went up

Images 2.0 would often ignore a "keep everything else the same" instruction and re-imagine the frame anyway. If 2.5 is genuinely better at leaving unmentioned things alone, then the prompt's job splits cleanly into two lists: what to change and what to preserve. Prompts that only describe the desired end state will still work, but they leave the model to guess which parts of the source were load-bearing.

Before:

Prompt
Make this photo look like it was taken at golden hour.

After:

Prompt
Edit the attached photo. CHANGE: the light to late golden hour, warm low sun from camera-left, long soft shadows. PRESERVE: the subject's face, pose, clothing, the framing and crop, the background buildings and their positions. MATCH: the new light must fall consistently on the subject and the background.

The second prompt is not longer for its own sake. Every clause is a decision the model would otherwise make for you. This is the CHANGE / PRESERVE / MATCH structure we have used since the GPT Image 2 era; 2.5 is the first model where we expect the PRESERVE list to be respected most of the time rather than some of the time.

Say how the reference should be used

"Better at preserving the subjects in reference photos" is only useful if the model knows the photo is a subject reference and not a style or composition reference. A photo of your friend can mean "put this person in the scene," "copy this lighting," or "copy this framing." State it:

Prompt
Reference image: use as SUBJECT IDENTITY only. Keep the face, hair, and build recognizable. Do not copy the reference's clothing, background, or lighting.

The Prompt workspace asks for this explicitly (style, subject identity, edit source, product, composition, or sketch layout) and writes the corresponding clause into the prompt.

Put text and layout in their own sentences

The layout improvements are the ones most likely to be over-read. "More accurate real-world content" and "complex layouts including transparent backgrounds" mean the model is more likely to honor a defined structure, not that it will invent one for you. For posters, slides, and infographics, name the exact text in quotes, give an order, and say what is largest:

Prompt
Exact text, verbatim: headline "OPEN STUDIO" (largest, top third), subhead "Saturday 14 Sept, 10:00-18:00" (below headline, half its size), footer "Free entry" (bottom-left, small). No other text anywhere in the image.

The last sentence matters. Unlabeled space invites invented copy.

Multi-turn: treat each edit as a new prompt

Better consistency across turns does not mean the model remembers your constraints. It means it is less likely to drift when you repeat them. In a long session, restate PRESERVE on every edit. Copying the previous turn's PRESERVE list forward costs you ten seconds and saves the face.

What did not change

The aspect ratio still needs to be stated. Vague quality adjectives still do nothing that a concrete constraint would not do better. And a prompt that contradicts itself ("minimal, with lots of detail") still produces a coin-flip. The Critique mode scores exactly these things: intent fidelity, contradictions, reference handling, edit preservation, text and layout, and efficiency.

What Depikt changed

  • The Prompt Builder now targets ChatGPT Images 2.5. Its intent stage extracts the format, reference use, aspect ratio, exact text, series count, and transparent background before writing a word of the prompt.
  • The Prompt Critic scores prompts against a 2.5 rubric, with a hard cap when an essential dimension such as edit preservation fails.
  • The 500-prompt Library stays what it is: the GPT Image 2 collection, kept as written. Remix any entry to get a version rewritten for 2.5. Reviewed Images 2.5 recipes sit alongside it in Library.

What to read next

Sources

Build a prompt that follows this guide.

Paste a rough idea. Get back a structured prompt.

Open Prompt

More in Images 2.5

Images 2.5
How to Prompt ChatGPT Images 2.5

Images 2.5 is better at doing exactly what you say and leaving alone what you do not mention. That changes where the effort in a prompt should go. Here is the six-part structure we use, the mistakes that still waste generations, and five rough ideas rewritten into prompts that ship.

Images 2.5
Precise Edits and Reference Images in ChatGPT Images 2.5

OpenAI's two biggest claims for Images 2.5 are better reference fidelity and edits that change only what you asked for. Both depend on the prompt saying what the reference is and what must stay. Here is how to write those prompts, with examples for identity, style, product, composition, and sketch references, and a routine for multi-turn edits.