Image generation

ChatGPT Images 2.5

by OpenAI

OpenAI's expected next image model.
What testing has shown, and what is still unconfirmed.

ChatGPT Images 2.5

Key features

None of the above is confirmed by OpenAI. The four expected changes come from reports of anonymous pre-release testing in August 2026; the last two describe GPT Image 2 behavior that an interim release would carry forward. Last reviewed September 2026.

Technical specifications

Not yet

No matching model id in OpenAI's SDKs, changelog, or docs as of September 2026.

2 checkpoints

mona-lisa-1 and luna-lisa-alpha, both in anonymous Arena battles during August 2026.

Up to 2K

GPT Image 2's output ceiling, the baseline any interim release would start from.

Up to 8

GPT Image 2 returns up to eight images per prompt with shared characters and style.

Use cases

A film festival poster headlined NORTHLIGHT with a full credit block in small type beneath it

Posters and packaging

Layouts where the headline and the small print both have to be readable. The work a typography release would change most.

A labelled cutaway diagram of an espresso machine with a small caption on every part

Infographics and diagrams

Charts and labeled process sheets, where a dozen small captions all have to survive at once rather than one clean headline.

The same transit poster laid out three times in English, Japanese and Arabic

Multilingual campaign sets

The same layout rendered across scripts. Non-Latin glyph accuracy is the axis with the most room left in the current model.

The same tea tin shown twice, cream paper label beside matte black, framing identical

Iterative product edits

A long back-and-forth on one image, where the third and fourth edits are where the frame usually starts drifting.

Six storyboard panels following the same red-jacketed courier in one painted style

Storyboards and sequences

Batches that have to read as one world, with the same character and props holding across every frame in the set.

A settings screen mockup on a laptop with correctly spelled interface labels

UI and screen mockups

Interface frames with correctly spelled button labels, an area GPT Image 2 was already noted for at launch.

Prompt examples

A wine bottle label with an estate name, vintage year and four lines of tasting notes

Wine label

A wine bottle label, estate name, vintage year, and four lines of tasting notes

Edit prompt
A national park trail map poster with six routes each labelled with a name and distance

Trail map

A national park trail map poster, six routes labelled with names and distances

Edit prompt
A boarding pass mockup with passenger, gate, seat and time set in a clean grid

Boarding pass

A boarding pass mockup, passenger, gate, seat and time legible in a clean grid

Edit prompt
A three-panel comic strip with hand-lettered speech balloons in ink and flat color

Comic strip

A three-panel comic strip with hand-lettered speech balloons, ink and flat color

Edit prompt
An award certificate with an ornate border, a recipient name and a date in script

Certificate

An award certificate, ornate border, name and date in calligraphic script

Edit prompt
The back panel of a cereal box showing a full nutrition table with legible figures

Nutrition panel

A cereal box back panel, a full nutrition table with every figure legible

Edit prompt

Simple pricing

Get started for free today, with the option to upgrade or cancel anytime.

Basic

$9/ month
billed as $0 per year

1100 monthly credits

1 user only

All models

Workflows

Standard

$24/ month
billed as $0 per year

3625 monthly credits

1 user only

All models

Workflows

Pro

$45/ month
billed as $0 per year

6350 shared monthly credits

1 user

+ up to 4 more at extra cost

All models

Workflows

Pro Max

$170/ month
billed as $0 per year

24650 shared monthly credits

1 user

+ up to 9 more at extra cost

All models

Workflows

Enterprise

For higher limits

Custom

pricing and billing terms

High-volume credits
Custom seat limits
All models
Workflows
Pricing Gradient

Free

For playing around

$0

forever free

Up to 20 credits
1 user only
Limited models
Workflows

FAQs

Is ChatGPT Images 2.5 real?
Not officially. As of September 2026 OpenAI has not announced a model by that name, and no matching id appears in the OpenAI API changelog, the documentation, or the ImageModel type in the openai-python and openai-node SDKs, which still end at gpt-image-2 alongside the chatgpt-image-latest alias. What does exist is evidence of pre-release testing: two anonymous checkpoints, mona-lisa-1 and luna-lisa-alpha, appeared in Arena battles during August 2026. This page collects what is known and updates when OpenAI announces something.
When is ChatGPT Images 2.5 expected?
There is no announced date. The one usable signal is the pattern from the last launch: three anonymous checkpoints ran on Arena around April 5 and 6, 2026, and GPT Image 2 shipped on April 21, roughly fifteen days later. Applying that gap to luna-lisa-alpha, which surfaced around August 19, points at early September 2026. OpenAI's DevDay on September 29, 2026 is the next scheduled stage moment. Both are inferences from precedent, not schedules.
Will it be called ChatGPT Images 2.5 or GPT Image 3?
Unresolved. The naming precedent cuts toward a point release: gpt-image-1.5 shipped in December 2025 as an interim quality bump between gpt-image-1 and gpt-image-2, so a 2.5 would fit the established rhythm. A larger jump would more likely carry a whole number. Community discussion has floated 2.5, 3, a mini variant, and a silent dated snapshot under the existing gpt-image-2 id. None of these appear in any OpenAI source yet.
What is expected to improve over GPT Image 2?
Three themes recur in reports from pre-release testing: finer typographic control at small sizes, more convincing texture under harsh lighting, and better consistency across a chain of edits to the same image. A more recent knowledge cutoff than GPT Image 2's December 2025 has also been mentioned. All four are qualitative, single-source, and unverified. Nothing here should be read as a specification.
What were mona-lisa-1 and luna-lisa-alpha?
Two anonymous model checkpoints that surfaced in LM Arena's blind battle mode during August 2026, on August 9 and around August 19. Neither appeared on a public leaderboard, and neither carried any OpenAI provenance: no watermark, no tokenizer match, no Arena attribution. Both were reported by the same account. The alpha suffix on the second suggests further candidates may run before anything ships. OpenAI has said nothing about either.
How reliable is the information on this page?
It separates two tiers. The verifiable parts are the GPT Image 2 launch record, the SDK and changelog contents, the gpt-image-1.5 precedent, and the DevDay date. The unverifiable parts are the capability reports, which trace to a single account posting about anonymous Arena checkpoints. Everything in the second tier is marked as reported or expected throughout. The page will be rewritten against the official announcement when there is one.