How to use ChatGPT Images 2.5: prompts, editing, and what to expect

How to use ChatGPT Images 2.5: prompts, editing, and what to expect

A plain guide to ChatGPT Images 2.5 (GPT Image 2.5): how to write your first prompt, get text right, build a layout, edit without breaking the image, and what the release is expected to change.

Try on Morphic

What is ChatGPT Images 2.5?

ChatGPT Images 2.5, also written GPT Image 2.5, is the name people are using for OpenAI's expected next image model. It has not been announced, and no model by that name exists in the OpenAI API today.

What does exist is a testing trail, and a prompt style that has carried across every release in this family. So most of this guide teaches you to get good images out of the current model, in order. What you learn here should transfer the day anything new lands. The last section covers the release itself.

The ChatGPT Images 2.5 prompt formula

Four habits do most of the work. Everything below is an example of one of them.

HabitWhat it looks likeWhat it fixes
Describe a picture, not keywordsA vintage travel poster of a harbour town at golden hourVague output from adjective lists
Quote the exact wordsThe sign reads "HARBOR LIGHT" in a bold serifMisspelled and invented text
Name the type hierarchyHeadline large at the top, four acts beneath in descending sizesFlat, unstructured layouts
One change per edit turnChange the label to matte black, keep everything else exactlyDrift across a chain of edits

Writing your first image prompt

Name the subject, the setting and the light in plain sentences. Words like "beautiful" or "4k" do almost nothing, because they do not tell the model what to put in the frame.

Weak promptStrong prompt
travel poster, beautiful, high quality, 4kA vintage travel poster of a harbour town at golden hour, warm flat color, the title across the top

Start there, look at what comes back, then change one thing at a time.

A vintage travel poster of a harbour town at golden hour with the title set across the top
One sentence of subject, setting and light. No stacked adjectives.

Getting text and layout right in an image

This is the most useful skill to build, and the one the expected release is said to improve most. Put the exact words in quotation marks and say where they sit. Keep the wording short, because long runs of small type degrade first.

A bakery menu board with six items and prices in tidy chalk-lettered columns
Six items and prices quoted in the prompt, so the model renders them instead of inventing them.

Once single lines work, name the parts of the layout separately instead of describing a mood. "A concert bill: headline act large at the top, four supporting acts stacked beneath in descending sizes, venue and date small at the base" gives the model a hierarchy to plan against.

A concert bill with the headline act large and supporting acts stacked in descending sizes
A named hierarchy: headline act, supporting acts in descending sizes, venue and date small at the base.

Editing an image and keeping a set consistent

Editing is where most people lose a good image. Change one thing per turn, and say what to leave alone.

Do not writeWrite instead
Make it darker and more premium, maybe move the logoChange the label to matte black, keep the lighting and framing exactly as they are
The courier again, different sceneThe same courier in a red jacket and grey scarf, now on a mountain pass

Watch for drift. The first two edits usually hold; by the third or fourth something you never asked about has moved. Go back to the last good version rather than correcting forward.

The same tea tin shown twice, cream paper label beside matte black, with framing and lighting unchanged
One change, everything else held. Only the label differs; lighting and framing are untouched.

For a set that has to read as one world, repeat the recurring details word for word. Where the model can generate a batch from one prompt, use it: a set produced together holds far better than six images made separately.

Six storyboard panels following the same red-jacketed courier in one consistent painted style
The same courier described word for word across all six frames.

ChatGPT Images 2.5 release date and expected features

Everything known about the release falls into two tiers, and the gap between them matters more than any single claim.

ClaimSourceWeight
GPT Image 2 shipped April 21, 2026OpenAI API docs and changelogVerifiable
gpt-image-1.5 shipped December 2025 as an interim releaseOpenAI announcement recordVerifiable
No 2.5 or 3 id in the SDKs, changelog or docsopenai-python and openai-nodeVerifiable
Two unnamed checkpoints ran in Arena, August 2026One X account, no Arena attributionSingle source
Sharper small text, better texture, steadier editsSame account, qualitativeSingle source

The checkpoints before GPT Image 2 appeared about fifteen days ahead of launch. Applied to the most recent sighting that points at early September 2026, with DevDay on September 29 the next natural moment. Neither is a schedule, and a checkpoint that runs in Arena does not have to ship.

To check whether it has launched, look for a new image id in the OpenAI API model list, a dated entry in the changelog, and a new value in the ImageModel type shipped with the SDKs. A real launch clears all three.

GPT Image 2 is on Morphic today and is the model any 2.5 release would be measured against. Building your prompts on it now means the work transfers on day one.

FAQs

Has OpenAI announced ChatGPT Images 2.5?
No. As of September 2026 there is no announcement, no changelog entry, and no matching model id in the OpenAI API or SDKs. The ImageModel type in openai-python and openai-node ends at gpt-image-2 alongside the chatgpt-image-latest alias. Everything circulating under the name traces back to anonymous checkpoints seen in Arena testing during August 2026.
What is the evidence that a new image model is coming?
Two anonymous checkpoints, mona-lisa-1 around August 9 and luna-lisa-alpha around August 19, 2026, appeared in LM Arena's blind battle mode and were reported by a single account. On its own that is weak. It draws attention because the pattern matches the previous launch, where three anonymous checkpoints ran on Arena around April 5 and 6 and GPT Image 2 shipped on April 21.
What is ChatGPT Images 2.5 expected to change?
Three themes recur in the reports: finer typographic control at small sizes, more convincing texture under harsh direct lighting, and better consistency across a chain of edits to one image. A newer knowledge cutoff than GPT Image 2's December 2025 has also been mentioned without a date. All of these are qualitative and unverified.
Do prompts written for GPT Image 2 carry over?
Historically yes. Every model in this family has kept the same prompt surface, with new releases raising the ceiling on what the same prompt returns rather than changing how you write it. Prompts built around explicit quoted text, stated type hierarchy, and one change per edit turn have survived each previous transition and are the safest thing to build now.
How do I get text to come out correctly in an AI image?
Put the exact words in quotation marks and name where they sit. Asking for a sign that reads HARBOR LIGHT in a bold serif across the top beats describing a sign with the shop name on it, because the model then has a string to render rather than a concept to invent. Keep the wording short, and ask for fewer words at larger sizes when accuracy matters most.
How can I check whether it has actually launched?
Three checks settle it in under a minute: look for a new id in the OpenAI API model list, check the API changelog for a dated entry, and check whether the ImageModel type in the current openai-python or openai-node release has gained a value. A model that has shipped appears in all three. Coverage that cites only an Arena screenshot has not cleared any of them.
Should I wait for it before starting a project?
No. The reported changes are refinements to typography, texture, and edit stability rather than a new capability class, and no date is confirmed. GPT Image 2 already handles the work these reports describe, at a lower quality ceiling. Building the prompts now means the work transfers on day one if a new model lands.