Direct answer: OpenAI released ChatGPT Images 2.5 on September 8, 2026. It is rolling out to ChatGPT, ChatGPT Work and Codex users on desktop, mobile and web. Developers can use two API models: gpt-image-2.5-flare for fast everyday generation and gpt-image-2.5-sunburst when editing precision and premium output matter more than speed.[1][2]
Quick start
- In ChatGPT: describe the image you want, use
@Sketchto draw a rough visual guide, or start from a Poster or Merch template.[1] - For fast API work: choose
gpt-image-2.5-flare.[1][2] - For detailed edits or polished campaign assets: choose
gpt-image-2.5-sunburst.[1][2] - For one-shot generation or editing: use the Images API.[2]
- For conversational, multi-turn image editing: use the Responses API.[2]
- Current token rates for both 2.5 models: $8 per million image-input tokens, $2 per million cached image-input tokens, $30 per million image-output tokens, $5 per million text-input tokens and $1.25 per million cached text-input tokens.[2]
What changed in ChatGPT Images 2.5?
OpenAI says the new release improves detail, natural lighting, texture, subject preservation and instruction-following during edits. It also says generation latency is up to 50% lower than Images 2.0.[1]
The most useful changes are practical rather than cosmetic:
- More focused edits: ask to change one product, background or line of copy while retaining the rest of the composition.[1]
- Better multi-turn consistency: approved details are more likely to survive later edits in the same conversation.[1]
- Sketch input: type
@Sketchin ChatGPT, draw a rough layout, then explain the desired finish.[1] - Templates: start from common formats such as posters, merchandise or product photos instead of a blank prompt.[1]
- Comment-based editing: place comments directly on an image to point at the part that needs changing.[1]
- Prompt sharing: include the prompt when sharing an image so another person can reuse the idea with their own details.[1]
- Transparent-background support: the models can handle more complex layouts, including transparent backgrounds.[1][2]
OpenAI says Images 2.5 is available across all ChatGPT tiers, but “available” does not mean every account receives identical limits. Your plan, workspace controls and rollout timing can still affect what you see.[1]
Flare vs Sunburst: which model should you choose?
| Model | Best for | Main trade-off |
|---|---|---|
gpt-image-2.5-flare |
Social posts, product experiences, visual search, prototypes and high-volume generation | Optimized for speed; use Sunburst when the finest edit control matters |
gpt-image-2.5-sunburst |
Polished product imagery, campaign creative and precision-heavy editing | Longer generation time |
OpenAI positions Flare as the default for most applications. Sunburst is the premium choice for workflows where tighter control across edits is worth waiting longer.[1]
A sensible production workflow is to prototype with Flare, approve composition and copy, then rerun only the final candidates with Sunburst. Do not assume one model is always cheaper per finished asset: both use the same token rates, but token consumption can differ by model, size and quality.[2]
How to use ChatGPT Images 2.5 without code
1. Start with the deliverable, not a vague style
Say what you need: a 16:9 blog hero, a square product post, a portrait event flyer or a transparent product cutout. Include the audience and where the asset will appear.
2. Lock the non-negotiable details
List exact text, subject, composition, colors, aspect ratio and items that must not change. For an edit, explicitly separate the requested change from the protected details.
Reusable prompt formula:
Create a [format] for [audience/use]. Show [subject] in [setting and composition]. Use [lighting, palette and style]. Include the exact text: “[copy]”. Keep [protected details] unchanged. Output [aspect ratio/background requirement]. Avoid [specific unwanted elements].
3. Use Sketch when position matters
Type @Sketch, draw the rough position of the headline, product and call-to-action, then describe the visual style. A basic box-and-arrow drawing is enough; OpenAI says the feature is intended for users who are not professional artists.[1]
4. Refine one variable at a time
Make a focused request such as “change only the background to warm grey” or “replace the price line while preserving the product, shadows, camera angle and all other text.” This makes it easier to detect unwanted drift.
5. Perform a final human check
Read every letter, verify product labels and logos, compare reference identities and inspect the entire frame. OpenAI’s prompting guide specifically recommends checking text accuracy, diagram relationships, product shapes, reference details, edit scope and real alpha transparency before use.[8]
GPT Image 2.5 API setup in Python
Install the current OpenAI Python package and place the API key in an environment variable rather than hard-coding it.
python -m pip install --upgrade openai
export OPENAI_API_KEY="your_api_key_here"
The Images API is the simplest route for a single generation. The official API returns base64 image data, which you decode and save.[2]
from openai import OpenAI
import base64
client = OpenAI()
result = client.images.generate(
model="gpt-image-2.5-flare",
prompt=(
"Create a clean 16:9 editorial illustration for a tutorial about "
"AI image editing. Dark navy background, cyan and violet accents, "
"large legible title 'IMAGE EDITING WORKFLOW', no logos."
),
quality="medium",
size="1536x1024",
)
image_bytes = base64.b64decode(result.data[0].b64_json)
with open("image-editing-workflow.png", "wb") as file:
file.write(image_bytes)
Use gpt-image-2.5-sunburst instead when the workflow depends on precision-heavy final output. The API guide says an organization may need to complete API Organization Verification before using GPT Image models.[2]
curl example
curl -X POST "https://api.openai.com/v1/images/generations" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2.5-flare",
"prompt": "A square product photo of a matte black reusable bottle on pale stone, soft side light, realistic shadow, no text",
"quality": "medium",
"size": "1024x1024"
}' \
| jq -r '.data[0].b64_json' \
| base64 --decode > product-photo.png
Images API or Responses API?
Choose the Images API for a single generation or edit from one prompt. Choose the Responses API when the image belongs in a conversation or multi-step application and users need iterative editing.[2]
The Responses API can preserve an image workflow through previous_response_id. It can also use action: "generate", action: "edit" or action: "auto"; forcing an edit without an image in context returns an error.[2]
GPT Image 2.5 pricing explained
Both Flare and Sunburst currently use the same token rates:[2]
| Usage type | Price per 1 million tokens |
|---|---|
| Image input | $8.00 |
| Cached image input | $2.00 |
| Image output | $30.00 |
| Text input | $5.00 |
| Cached text input | $1.25 |
Your total request cost is the sum of text input, reference-image input and generated image-output tokens. The exact image cost varies with model, quality and dimensions, so use the usage field from real requests rather than budgeting from a single headline number.[2]
Quality options for both 2.5 models are low, medium, high, xhigh and max. OpenAI’s calculator also warns that auto can choose a different output profile, so set explicit size and quality values when cost predictability matters.[2]
If streaming partial images, each partial image adds 100 image-output tokens. Responses API workflows also incur the selected mainline model’s token usage in addition to image-generation costs.[2]
10 copy-paste prompts to test
Product photo
Create a square ecommerce product photograph of [product] centered on [surface]. Use [lighting] and a [color] background. Preserve the exact product shape, label and logo. Realistic shadow, no extra objects, no invented text.
Targeted product edit
Change only the background to [new background]. Preserve the product, label text, camera angle, crop, reflections, shadows and dimensions exactly.
Blog featured image
Create a clean 16:9 editorial illustration for an article titled “[title]”. Use a simple visual metaphor: [metaphor]. Leave generous empty space around the edges. No logos, no photorealistic public figures and no text except the exact title.
Event flyer
Design a portrait event flyer. Exact headline: “[headline]”. Exact details: “[date, time, venue]”. Visual hierarchy: headline first, date second, venue third. Use [palette/style]. Keep all wording legible and do not add extra copy.
Transparent product cutout
Isolate [object] from the reference image. Preserve its shape, color, label and surface texture. Remove every background element and return a true transparent background with a clean alpha edge.
Social carousel cover
Create a 4:5 social carousel cover about “[topic]”. Large exact title, one simple supporting icon and strong mobile contrast. Keep the bottom 12% clear for interface overlays.
Infographic
Create a vertical infographic explaining [process] in exactly [number] numbered steps. Use the exact step labels supplied below. Keep arrows and relationships logically correct. Do not add statistics or claims.
Room sketch conversion
Turn this rough room sketch into a realistic interior. Preserve wall openings, furniture positions and camera viewpoint. Apply [style], [materials] and [lighting]. Do not add doors or windows.
Consistent character edit
Keep the character’s face, hair, age, clothing, body proportions and pose unchanged. Change only [requested detail]. Match the existing lighting and lens perspective.
Ad variation
Create three visual variations of the approved ad while preserving the product, offer wording, logo placement and brand colors. Vary only the background composition and supporting props.
Safety, provenance and commercial review
OpenAI says Images 2.5 uses prompt and image checks, C2PA metadata and invisible watermarking. Its system card also says increased realism can create more convincing deepfakes if safeguards fail, so political, sexual or otherwise sensitive depictions of real people require particular care.[1][7]
Do not treat a generated image as proof of an event. For commercial use, review rights to every uploaded reference, trademark, face and piece of copy; obtain human approval before publishing customer-facing assets. A generation completing successfully is not a legal clearance or an accuracy guarantee.
Practical launch checklist
- [ ] Choose Flare for fast iteration or Sunburst for precision-heavy output.[1][2]
- [ ] Set explicit size and quality for predictable testing.[2]
- [ ] Keep API keys out of scripts and repositories.
- [ ] State exact text and protected details in the prompt.[8]
- [ ] Change one variable per edit.
- [ ] Check spelling, labels, anatomy, edge artifacts and transparency.[8]
- [ ] Log model, prompt, size, quality, usage and output version.
- [ ] Confirm rights to all reference assets.
- [ ] Label editorial or AI-generated imagery where context could mislead.
- [ ] Keep a human approval step for ads, news, healthcare, finance and public-figure content.
Frequently asked questions
Is ChatGPT Images 2.5 free?
OpenAI says Images 2.5 is rolling out across all ChatGPT tiers. Usage limits can still depend on the account, plan and workspace. API usage is separately metered at the published token rates.[1][2]
How do I open Sketch?
Type @Sketch in ChatGPT, draw a rough visual guide, then add a text description of the desired image and style.[1]
Is Flare or Sunburst better?
Flare is the practical default for speed and everyday generation. Sunburst is intended for premium workflows requiring tighter control and more precise edits, but it takes longer.[1][2]
Can I use GPT Image 2.5 through the API?
Yes. Use gpt-image-2.5-flare or gpt-image-2.5-sunburst with the Images API, or specify one of them inside the Responses API image-generation tool.[2]
Does the API return an image URL?
The official examples for these models read b64_json, decode the base64 data and save it as an image file.[2]
Does Images 2.5 support transparent backgrounds?
Yes. OpenAI lists transparent-background handling among the improved layout capabilities, and its API documentation supports customizable background and output settings.[1][2]
Bottom line
ChatGPT Images 2.5 is most useful as an editing and production-workflow upgrade: faster iteration, stronger subject preservation, focused comments, rough sketches and two API tiers. Start with Flare, move final precision-heavy assets to Sunburst, and measure actual usage and output quality on your own representative prompts before migrating a production pipeline.[1][2]
Sources
[1] https://openai.com/index/introducing-chatgpt-images-2-5 — Introducing ChatGPT Images 2.5 | OpenAI
[2] https://developers.openai.com/api/docs/guides/image-generation — Image generation | OpenAI API
[7] https://deploymentsafety.openai.com/chatgpt-images-2-5 — ChatGPT Images 2.5 System Card
[8] https://developers.openai.com/api/docs/guides/image-prompting — Image prompting | OpenAI API