AI NewsSep 9, 2026Words 1495Read time≈ 4 min

OpenAI Launches ChatGPT Images 2.5 With Flare and Sunburst API Models

OpenAI is rolling out ChatGPT Images 2.5 and launching Flare and Sunburst API models, with faster generation and tighter editing.

AIOpenAI

Contents · 12

1. Images 2.5 Focuses on Repeatable Editing
2. Sketches, Comments, and Templates Change the Interface
3. Flare and Sunburst Give Developers Two Deployment Options
4. The Two Models Share Pricing but Not Necessarily Per-Image Cost
5. Higher Realism Comes With the Existing Provenance Stack
Frequently Asked Questions
When was ChatGPT Images 2.5 released?
Who can access Images 2.5?
Which API model should developers choose?
Is Sunburst also 50% faster than GPT-Image-2?
Are templates available in ChatGPT Work?
Sources

OpenAI released ChatGPT Images 2.5 on September 8, 2026, less than five months after introducing Images 2.0. The company says the new version cuts image-generation latency by up to 50% while improving reference-photo fidelity, targeted edits, and consistency across successive revisions.

The rollout covers all tiers of ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and the web. OpenAI also launched two developer models: GPT-Image-2.5 Flare, positioned as the faster default for everyday applications, and GPT-Image-2.5 Sunburst, designed for detailed creative and editing work where precision takes priority over generation time.

The release is relevant at unusually large scale. OpenAI says users now create more than 3 billion images each week across ChatGPT Images and its GPT-Image API models.

1. Images 2.5 Focuses on Repeatable Editing

OpenAI’s main claims concern what happens after the first image is generated. Images 2.5 is intended to change a requested element without unnecessarily redrawing the surrounding subject, composition, background, or visual treatment.

That behavior addresses a practical weakness of earlier image generators: even a small revision could alter a person’s face, a product’s shape, the placement of text, or other details the user wanted to preserve. OpenAI says Images 2.5 is better at retaining recognizable subjects from reference photographs and maintaining previous changes during multi-turn editing.

The company also reports more natural lighting, richer textures, more accurate use of real-world information, and better handling of complex layouts. OpenAI specifically identifies transparent backgrounds, branded asset series, presentation graphics, user-interface concepts, and product imagery as workloads expected to benefit.

These are primarily vendor claims rather than results from a published comparative benchmark. Axios conducted early hands-on tests involving a cat photograph, a tattoo design, and a logo. Its reporter found that the system preserved the cat’s likeness and supported detailed, visually targeted revisions, offering limited independent support for the editing claims but not a quantitative measurement of the reported latency reduction.

2. Sketches, Comments, and Templates Change the Interface

Images 2.5 arrives with four creation and sharing features inside ChatGPT: Sketch, templates, image comments, and shareable prompts.

Sketch lets a user draw a rough visual reference directly in ChatGPT and combine it with a written description. Typing @Sketch invokes the feature, although OpenAI’s release notes describe the initial mobile flow as typing @ and selecting Sketch. A drawing can communicate spatial information—such as the outline of clothing, the arrangement of a room, or the location of objects—that may be cumbersome to specify through text alone.

Templates provide predefined starting points for formats such as posters, merchandise, flyers, and product photographs. Users supply the message, desired style, and other details instead of constructing the full prompt from an empty input field. ChatGPT may ask follow-up questions before generating the image.

Comments offer a more localized editing workflow. On supported mobile interfaces, users can open an image, identify the area they want changed, and attach an instruction to that location. Prompt sharing allows the creator of an image to include the underlying prompt so another user can adapt it with different photographs or details.

The rollout has product-specific qualifications. OpenAI’s release notes say templates are not yet available in Work mode, despite the underlying Images 2.5 model reaching ChatGPT Work. Existing image-generation usage limits also remain unchanged. Some interface elements, including the new loading experience, vary by platform and rollout status.

3. Flare and Sunburst Give Developers Two Deployment Options

The API release separates the Images 2.5 family into two named models rather than presenting one model with only adjustable quality settings.

GPT-Image-2.5 Flare is OpenAI’s recommended default for most applications. The company describes it as suitable for social content, visual search, product experiences, rapid prototyping, and high-volume generation. OpenAI claims Flare produces higher-quality images than GPT-Image-2 with 50% lower latency.

GPT-Image-2.5 Sunburst is the more capable option for image generation and editing. It targets campaign creative, polished product imagery, and other workflows where preserving details across edits matters more than completion time. OpenAI does not publish a numerical latency comparison for Sunburst and explicitly warns that it has longer generation times.

Developers can select the models using the identifiers gpt-image-2.5-flare and gpt-image-2.5-sunburst. Dated snapshots, gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08, are available for applications that need to pin behavior to the launch version.

Both models accept text and image inputs and return images. They can be used directly through the Image API for standalone generations or edits. They can also be selected inside the Responses API image-generation tool, which is the more appropriate interface for conversational workflows and repeated edits that retain prior context.

The models add xhigh and max above the existing low, medium, and high quality settings; auto remains available. They support recommended square, landscape, and portrait sizes as well as custom dimensions. Custom width and height must be multiples of 16, stay between 1:3 and 3:1 in aspect ratio, contain between 655,360 and 8,294,400 pixels, and use no edge longer than 3,840 pixels. Resolutions above 2,560 by 1,440 remain experimental.

Developers may need to complete API organization verification before using GPT-Image models. Output format, compression, quality, dimensions, and transparent or opaque backgrounds can be configured through the API.

OpenAI lists the same token rates for Flare and Sunburst: $5 per million text-input tokens, $1.25 per million cached text-input tokens, $8 per million image-input tokens, $2 per million cached image-input tokens, and $30 per million image-output tokens.

Identical token rates do not guarantee that the models cost the same for a completed image. Token consumption can vary by model, resolution, and quality setting. Applications using the Responses API must also pay for tokens consumed by the main language model orchestrating the image-generation tool. Streaming partial images add 100 image-output tokens for each partial image returned.

There is also a documentation inconsistency developers should note. OpenAI’s numeric pricing table lists GPT-Image-2 at $2.50 per million text-input tokens, $4 per million image-input tokens, and $15 per million image-output tokens—half the corresponding GPT-Image-2.5 rates. The Flare and Sunburst model pages nevertheless contain a prose note stating that their token rates “match GPT Image 2.”

The numeric prices for the 2.5 models agree between their individual model pages, the image-generation guide, and the central pricing table. The conflicting comparison with GPT-Image-2 appears to be a documentation error, but OpenAI had not resolved it at publication time. Developers migrating production workloads should rely on recorded API usage and billing data rather than assuming that Flare’s lower latency also means a lower cost per image.

5. Higher Realism Comes With the Existing Provenance Stack

OpenAI says Images 2.5 continues to attach C2PA metadata and invisible watermarks to generated images. Its safety system checks requests and images both before generation and before the final output is shown.

The system card identifies heightened realism as a potential source of more convincing political, sexual, and other sensitive deepfakes. OpenAI describes upstream policy checks, moderation of combined text-and-image inputs, and a downstream monitor that evaluates the generated image.

On OpenAI’s adversarial safety test set, the percentage of prompts resulting in an unsafe image being presented was 1.09% for Sunburst and 1.41% for Flare, compared with 1.64% for Images 2.0. These figures do not estimate the frequency of unsafe requests or outputs in normal use: the evaluation deliberately used prompts designed to elicit policy violations.

OpenAI further cautions that automated labels can contain errors and that category sample sizes affect statistical precision. The system card reports that none of the differences in the “unsafe presented” outcome reached its stated significance threshold. The results therefore document the operation of the safety stack, but do not establish a statistically significant overall safety improvement over Images 2.0.

Frequently Asked Questions

When was ChatGPT Images 2.5 released?

OpenAI announced and began rolling it out on September 8, 2026.

Who can access Images 2.5?

It is rolling out across all tiers of ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and the web. Individual interface features may arrive at different times.

Which API model should developers choose?

OpenAI recommends Flare for fast, everyday, or high-volume generation. Sunburst is intended for detailed workflows where editing precision matters more than latency.

Is Sunburst also 50% faster than GPT-Image-2?

OpenAI gives the 50% lower-latency comparison specifically for Flare. It does not publish a comparable Sunburst figure and says Sunburst takes longer to generate images.

Are templates available in ChatGPT Work?

Not yet. OpenAI says the Images 2.5 model is available in Work, but templates were unavailable in Work mode at launch.