ChatGPT Images 2.5: Flare vs Sunburst, Pricing, and Creative Workflows

Changing a product's background should not mean changing its label. Asking for a different headline should not mean rebuilding an approved composition. For a designer or developer, the useful question is not simply whether an image model can make something impressive. It is whether the next revision will keep the work worth keeping.

OpenAI released ChatGPT Images 2.5 on September 8, 2026, alongside two API models: GPT Image 2.5 Flare and GPT-Image-2.5-Sunburst. The developer release adds a choice between speed-oriented generation and precision-oriented creative work. 2 10

This guide examines the release from a production perspective: what changed, which model to test, what the pricing actually means, and how to move from a promising generation to usable campaign assets.

Research note: This is a source-based launch analysis, not an Oppye hands-on benchmark. Official specifications, outside users' observations, and our workflow recommendations are identified separately. Availability, documentation, and pricing reflect the verification date above.

Table of contents

• Flare vs Sunburst: which model should you use?
• What changes in the creative workflow?
• Image sizes, quality settings, and transparency
• GPT Image 2.5 pricing: what does an image actually cost?
• What early users are reporting
• How to prompt GPT Image 2.5 for production work
• How developers can use GPT Image 2.5
• From a generated image to a campaign-ready set
• Frequently asked questions

What is ChatGPT Images 2.5?

ChatGPT Images 2.5 is OpenAI's September 2026 update to image generation and editing in ChatGPT. OpenAI describes improvements in reference-image fidelity, selective edits, and consistency through successive revisions, with generation latency reduced by up to 50% against Images 2.0. That speed figure is OpenAI's claim, not a guaranteed result for every request. 1

The product and API names describe different layers. ChatGPT Images 2.5 is the experience people use in ChatGPT. gpt-image-2.5-flare and gpt-image-2.5-sunburst are explicit model identifiers developers can select in API integrations. Both support image generation and editing. 2

That distinction matters when reading reviews. A screenshot from ChatGPT does not, by itself, establish which API model or quality setting produced it. Treat claims about a specific model differently from impressions of the overall ChatGPT experience.

Flare vs Sunburst: which model should you use?

Use the Flare model when response time is the main constraint; test Sunburst when fine-grained editing and demanding visual requirements matter more. Both accept text and images and return images. They share a token rate card, but differ in their intended speed-quality trade-off. 3 4

The model positioning comes from OpenAI's prompting guide; the proposed test situations are our recommendations. The new xhigh and max settings are documented in the API changelog. 5 2

There is a useful nuance in the launch material. OpenAI's announcement describes Flare quality improvements, while its technical prompting guide characterizes Flare's image quality as comparable to GPT Image 2 and Sunburst's as higher. Do not turn that into a blanket promise that every Flare output will outperform its predecessor. 1 5

Our recommendation is to define an acceptance standard before choosing a winner. A concept image may pass when its composition and mood are right. An approved product visual may fail because a single character on the packaging changed.

For an existing workflow that already works, compare Flare against the current result. For a difficult brief that keeps failing, establish whether Sunburst solves the problem, then see whether Flare can meet the same standard. Optimize for the least expensive acceptable outcome, including review and correction time, rather than the most impressive model name.

What changes in the creative workflow?

The practical change is more control over how an image is revised, alongside new ways to communicate a visual brief. OpenAI highlights better preservation of subjects and surrounding details, plus stronger continuity through successive edits. These are improvements in reliability, not a promise of pixel-perfect preservation. 1

Local edits need a preservation brief

For production work, the instruction has two parts: what to change and what must remain unchanged. A background replacement can be successful aesthetically and still fail commercially if it alters the product's shape.

Our suggested review question is simple: Did the requested edit happen without breaking an approved detail elsewhere? Compare the whole image, not just the area you asked the model to modify.

Sketch and templates make the starting brief more concrete

ChatGPT's Sketch feature lets users draw a reference directly in the interface, using @Sketch, and add written instructions. Templates provide guided starting points for formats such as posters and logos. They help specify an image; they are not evidence that the output is an editable design document. 6

A useful agency application would be sketching the composition before generation: product on one side, headline space on the other, and a clear footer. That conveys hierarchy without spending a paragraph describing coordinates.

Comments and sharing support directed revision

The release adds comments placed on images and the ability to share the prompt behind a result. These features can make a brief or correction easier to communicate between collaborators. 1

We would still save approved versions separately. A shared prompt is a starting instruction, not a guarantee that a colleague will recreate the same pixels. Keep the reference, chosen output, and revision notes together so the team knows what was actually approved.

Image sizes, quality settings, and transparency

GPT Image 2.5 accepts custom resolutions subject to documented limits. A generation size is not automatically a valid ad export size. 7

For transparent output, explicitly request background="transparent" and choose PNG or WebP. JPEG is not the transparent-output option. The editing API also supports multiple reference images, with up to 16 documented inputs. 16

Calculated from those limits: 1080 × 1350 is not a valid native request, whereas 1536 × 1920 preserves its 4:5 ratio and can be downscaled. A 728 × 90 banner exceeds the permitted aspect ratio. These are mathematical implications of the published constraints, not additional restrictions. 7

Our preferred workflow is to separate composition from final delivery. Generate at a valid working size, inspect the content, and perform the exact final sizing and file preparation in a controlled export step. Do not rely on a prompt alone to deliver a compliant file.

GPT Image 2.5 pricing: what does an image actually cost?

Flare and Sunburst have identical published token rates, not a universal fixed price per image. OpenAI's standard API pricing distinguishes text input, image input, cached input, and image output. All figures below are USD per one million tokens, verified September 10, 2026. 8

An output estimate is not a complete request quote

OpenAI's guide now offers a 2.5-specific output-cost estimator. It excludes inputs; use the correct model selection and compare estimates with response usage. The old calculator is not interchangeable. 7

ChatGPT and the API have separate billing systems. A paid ChatGPT subscription does not pay for ordinary API requests. 9

Responses API charges include the coordinating model as well as image generation. 7

A worked cost example

Consider a hypothetical edit request, not a measured Flare or Sunburst benchmark, reporting 200 uncached text-input tokens, 1,500 uncached image-input tokens, and 1,800 image-output tokens.

Applying the published rates gives:

Text input:    200 / 1,000,000 x $5  = $0.001
Image input: 1,500 / 1,000,000 x $8  = $0.012
Image output:1,800 / 1,000,000 x $30 = $0.054
Total image-model request cost:       $0.067

The calculation is illustrative. It does not assert those token counts for a particular resolution or quality setting. Apply cached rates only to the cached portion of input; do not count that portion again at the uncached rate. 8

Budget for accepted assets, not only successful requests

For planning, we recommend this measure:

Cost per accepted asset =
all attributable generation and editing charges / accepted assets

Suppose a workflow averages 1.4 equally priced attempts per accepted asset. Using the hypothetical $0.067 request above, that becomes $0.0938 per accepted asset, or $93.80 for 1,000 accepted assets, before other costs.

This is why a slower model might still make business sense if it reduces correction rounds. Conversely, paying for a higher quality setting has little value when the first, faster result already meets the brief. Neither outcome should be assumed before testing.

What early users are reporting

Early feedback points to noticeable speed gains for some users, useful new controls, and continuing editing failures. It does not establish a broad performance consensus. The release was only two days old at this article's verification date; the examples below are observations, not a controlled comparison.

Developer discussion: cost visibility and revision behavior

In OpenAI's launch thread, developers questioned how unchanged token rates would translate into actual usage. One contributor also described a ChatGPT preview being replaced by a revised result they did not always prefer. Another shared Flare and Sunburst examples of an illustrated command-line reference poster. These are useful leads for testing cost, version retention, and dense layouts, not universal findings. 10

Read those launch-day documentation concerns alongside the more recent pricing information above, rather than treating every early complaint as a permanent limitation.

Community reactions: positive, mixed, and sometimes uncertain

An early Reddit discussion includes reports of faster generation and better facial expressions, alongside users who saw little difference or were unsure whether the rollout had reached them. One participant described a template producing the clarifying questions inside the poster rather than asking them first. These accounts are unverified, self-selected anecdotes without consistent settings. 11

Outside hands-on examples: useful controls, imperfect preservation

The Verge's Jay Peters reported turning a mouse-drawn cat sketch into a realistic image and using an image comment to change eye color. That is a concrete demonstration of the new interaction tools, not a technical benchmark of Flare against Sunburst. 12

A separate Promptsref review reported a more cautionary product-editing example: changing a bottle cap's color also changed its supporting stone stand. Other edits in the same small set worked better. Importantly, the reviewer disclosed that the underlying model identifier was not verified and that there was no matched Images 2.0 comparison set. The example supports inspecting surrounding details; it does not establish a failure rate for either API model. 13

The practical lesson is not that the launch claims are disproved. It is that better editing and dependable editing on your specific assets are different questions. Community examples help decide what to test, not what to promise a client.

How to prompt GPT Image 2.5 for production work

A production prompt should specify the deliverable, assign roles to reference images, identify the intended change, and protect the details that matter. OpenAI recommends explicit composition and constraints, narrow follow-up edits, and inspection of the result. 5

The examples below are original suggested briefs. They were not executed as tests for this article.

Example 1: replace the setting, preserve the product

Use the supplied product photograph as the product reference.
Create a landscape campaign image with the product on the right and
uncluttered space on the left for headline placement later.

Change only the environment: place the product on a matte stone surface
with soft daylight from the left and a quiet, warm-gray background.

Preserve the bottle silhouette, cap, proportions, label wording,
lettering, and product colors. Do not invent packaging details,
claims, certification marks, extra products, or decorative text.

Keep the result photographic rather than glossy or heavily retouched.
Do not add the headline or logo; those will be placed separately.

This brief separates photographic generation from final typography. Our suggested acceptance check is to compare the label, cap, silhouette, and colors against the supplied photograph before considering whether the new background looks attractive.

Where product fidelity is non-negotiable, consider compositing the approved product cutout over a generated background instead of asking a model to reconstruct every packaging detail.

Example 2: a narrow revision to an approved creative

Edit the supplied approved creative.
Replace only the headline with this exact text:
"LESS SETUP. MORE MAKING."

Keep its existing location, alignment, approximate visual weight,
and contrast. Preserve the product, background, logo, CTA, and all
other wording. Do not add a badge, new offer, extra copy, or new object.

Return a revised version, not a new creative direction.

This is a clear test of selective editing. Review punctuation, line breaks, and the rest of the image. A compelling result should still be rejected when it silently changes a required detail.

Practical applications beyond a single banner

For e-commerce, a useful pilot is adapting one approved product photograph to several seasonal settings. For agency work, test a sequence of small revisions rather than judging only the first output. For educational visuals, supply the facts and evaluate labels and relationships, not merely the illustration style.

OpenAI's prompting examples include diagram translation, product isolation, reference compositing, and multi-turn revisions. Those establish intended workflows, not a guarantee of factual accuracy or exact brand preservation. 5

How developers can use GPT Image 2.5

Use the Image API for direct generation and editing, or the Responses API when images belong inside a conversation or multi-step flow. With Responses, the image model is selected in the image_generation tool configuration; a supported coordinating model occupies the top-level model field. 14

For reproducible comparisons, record the actual model identifier. OpenAI lists dated snapshots including gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08. A fixed snapshot helps separate model changes from changes in your own prompts or settings. 3 4

A minimal generation request with usage logging

Install the current official openai Python package and set OPENAI_API_KEY in the environment. This illustrative example makes one billable request when run; it was not executed against the API for this article. It uses the documented generation method and response format. 15

import base64
import os
from pathlib import Path

from openai import OpenAI, OpenAIError

if not os.environ.get("OPENAI_API_KEY"):
    raise SystemExit("Set OPENAI_API_KEY before running this example.")

client = OpenAI(timeout=180.0, max_retries=0)

try:
    result = client.images.generate(
        model="gpt-image-2.5-flare-2026-09-08",
        prompt=(
            "Create a photographic studio background for a product campaign. "
            "Use matte stone, soft daylight, and clear space on the left. "
            "No products, people, text, or logos."
        ),
        size="1536x1024",
        quality="medium",
        output_format="png",
        n=1,
    )
    image = result.data[0] if result.data else None
    if not image or not image.b64_json:
        raise ValueError("The API returned no image data.")

    output = Path("gpt-image-2-5-example.png")
    output.write_bytes(base64.b64decode(image.b64_json, validate=True))
    print(f"Saved {output}")
    if result.usage is not None:
        print(result.usage.model_dump_json(indent=2))
except (OpenAIError, ValueError, OSError) as exc:
    raise SystemExit(f"Generation failed: {exc}") from exc

Keep API keys on the server, not in browser-delivered code. For a pilot, we recommend recording the prompt version, references, model, quality, dimensions, elapsed time, returned usage, and acceptance decision. Never treat HTTP success as proof that the image passed creative review.

Also check account limits before planning throughput. At verification, both model pages listed Tier 1 at 100,000 tokens and five images per minute, with higher limits at higher tiers. Launch-thread observations that limits were missing should not override the current documentation. 3 4

From a generated image to a campaign-ready set

An improved image model addresses generation and editing; campaign delivery still requires decisions about layout, approval, sizing, and exports. We recommend evaluating those as separate stages rather than treating a good-looking image as a finished campaign.

Start with a small collection of real briefs: a packaging-heavy product, a visual with required copy, a reference-led composition, and a sequence of targeted edits. Keep prompts and inputs consistent across models. Repeat enough runs to see whether the result is dependable, then compare first-pass acceptance, correction rounds, latency, and total cost.

Define rejection rules before reviewing the images. For example, an incorrect price, changed product geometry, or altered logo can be an automatic failure regardless of how polished the composition looks. This keeps aesthetic preference from hiding production mistakes.

Once a master creative is approved, the next question is how to adapt it without losing the approved idea. Oppye's Smart Resize is designed for that stage: it re-composes a creative for different formats, brings the set onto a whiteboard for comparison, supports individual AI edits, and prepares exact-dimension exports. These are Oppye workflow features, not capabilities being attributed to the OpenAI models. 18

The sensible handoff is therefore: create or refine the master, approve it, then adapt and review the campaign set.

For the next production step, see how to turn one ad into every size your campaign needs and the Performance Max image-specification checklist.

What still needs human review?

OpenAI still flags limitations in text placement, consistency, and layout control. Make these acceptance tests rather than assumptions. 7

Our recommendation is to preserve an approved reference and inspect each revision against it. Check all mandatory wording at final display size, not only at a comfortable zoom level. Keep important commercial claims and prices in a separately verifiable source of truth.

Provenance is another distinct check. OpenAI's documentation distinguishes signed C2PA metadata from embedded watermark signals and from visible labels. It also warns that a missing detection signal is not proof that media was made without AI. Editing or conversion can affect provenance signals. 17

Keep original files and an edit history where practical. Do not treat the presence of provenance metadata as a replacement for checking the content itself or reviewing the requirements of the intended publishing platform.

Frequently asked questions

When was ChatGPT Images 2.5 released?

OpenAI announced ChatGPT Images 2.5 on September 8, 2026. Its API changelog records Flare and Sunburst on the same date. This article was verified on September 10, 2026. 1 2

Is the Flare model cheaper than Sunburst?

Not at the published token-rate level: the two models share the same rates. Compare actual usage and the number of attempts needed to obtain an acceptable result before calling either cheaper for your workload. 8

Can GPT Image 2.5 create transparent backgrounds?

Yes. Request a transparent background and use PNG or WebP output. Inspect the exported file rather than assuming that a transparent-looking preview proves the file has the expected transparency. 16

Is ChatGPT Images 2.5 available to free users?

OpenAI's help documentation says ChatGPT Images is available across tiers. That does not mean every advanced experience is identical across plans: the help page lists separate availability for images with thinking. API billing remains separate from ChatGPT subscriptions. 6 9

Does Sunburst eliminate the need for a designer?

No such conclusion follows from the release. Sunburst is positioned for precise image creation and revision. Choosing the concept, checking brand accuracy, approving claims, and preparing a coherent campaign are separate responsibilities. 4

The production takeaway

The right adoption question is not whether ChatGPT Images 2.5 can produce an impressive demo. It is whether your team can reach an approved result with less waiting and less rework.

Set an acceptance standard, test both models where it matters, log the complete cost, and keep final delivery separate from generation. For AI image editing in 2026, the valuable improvement is not simply more detail. It is greater confidence that the next edit will preserve the work you already approved.

Related reading

• 10 AI Design Tools Every Graphic Designer Needs in 2026 — where a general-purpose image model sits alongside Photoshop, Illustrator, Figma, Ideogram, and the rest of a working stack.
• How to Turn One Ad Into Every Size Your Campaign Needs — current specs for Meta, Google Display, LinkedIn, and Performance Max, and the four aspect-ratio families that cover almost everything.
• Google Performance Max Banner Sizes (2026 Guide) — image and logo dimensions, the center-80% safe zone, and a production checklist.
• Do AI-Generated Ads Need a Label in the EU? — what Article 50 asks of advertisers, when an ad counts as a deepfake, and how Google, Meta, and TikTok treat AI-generated creative.
• Oppye documentation — supported platforms and sizes, how credits work, and how brand elements are carried through a re-composition.

Sources and verification

Sources were checked on September 10, 2026. Official documentation supports product and API facts; forum posts and outside demonstrations are cited only as attributed observations. The prompt examples and budgeting scenarios are original illustrations, not reported benchmark results.

1. OpenAI: Introducing ChatGPT Images 2.5. Launch announcement, September 8, 2026.
2. OpenAI API changelog. September 8 entries for the two image models.
3. GPT Image 2.5 Flare model documentation. Identifier, positioning, pricing, snapshots, and rate limits.
4. GPT Image 2.5 Sunburst model documentation. Identifier, positioning, pricing, snapshots, and rate limits.
5. OpenAI image-prompting guide. Model selection, references, editing patterns, and evaluation guidance.
6. OpenAI Help Center: Images in ChatGPT. Interface features and plan availability.
7. OpenAI image-generation guide. Size constraints, output-cost estimator, billing considerations, and limitations.
8. OpenAI API pricing. Standard image and text token rates.
9. OpenAI Help Center: Managing billing for ChatGPT and the API platform. Separate billing systems.
10. OpenAI Developer Community: Introducing GPT Images 2.5 in the API and ChatGPT. Launch discussion and individual observations.
11. Reddit: GPT Images 2.5 Discussion. Early, unverified user feedback.
12. The Verge: ChatGPT Sketch turns your bad drawings into detailed AI images. Jay Peters' firsthand Sketch and editing observations, September 8, 2026.
13. Promptsref: GPT Image 2.5 Review: Good Images, Uneven Edits. Small external example set; underlying model identifier not verified by its author.
14. OpenAI: Image generation with the Responses API. Tool configuration and conversation-based usage.
15. OpenAI API reference: Create image. Generation request and response fields.
16. OpenAI API reference: Create image edit. Editing inputs and output configuration.
17. OpenAI: Content provenance. Metadata, watermark signals, and verification limitations.
18. Oppye documentation. Smart Resize, whiteboard review, editing, and exports.