On September 8, 2026, OpenAI released ChatGPT Images 2.5, the successor to Images 2.0, along with two new API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. If you follow this site, you know we spend most of our time watching new image models appear and sorting what is verified from what is rumored. This launch is about as verified as it gets: an official announcement, published API docs, and same-day availability across ChatGPT tiers.
So here is the practical rundown: what GPT Image 2.5 changes, how Flare and Sunburst split the work, what happened to pricing (nothing), and where to try the model if you never want to write a line of API code.
What GPT Image 2.5 actually is
The name covers two things:
- ChatGPT Images 2.5, the image feature inside ChatGPT, rolling out to all tiers on desktop, mobile, and web.
- Two API models,
gpt-image-2.5-flareandgpt-image-2.5-sunburst, available to developers at launch.
OpenAI's headline claims for the release are sharper details, generation latency reduced by up to 50% compared with Images 2.0, more precise editing, and better preservation of subjects from reference photos. Those are vendor claims, not numbers we have measured ourselves, but the direction matches the first hands-on reports: cleaner textures, more natural lighting, and edits that change only what you asked for instead of re-rendering the whole scene.
A few capabilities matter most for everyday work:
- Reference-image fidelity. Give it photos of a subject and it holds likeness across styles and settings more reliably than GPT Image 2 did.
- Precision editing. You can ask for one element to change, such as a background, a product color, or a piece of text, and keep the rest of the composition intact.
- Multi-turn consistency. Earlier edits survive later edits, so iterating no longer means starting over every third step.
Flare vs. Sunburst: the two GPT Image 2.5 models
This is the decision most people hit first. The short version:
| GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst | |
|---|---|---|
| Positioning | Default for most work | Extra precision tier |
| Speed | Faster; latency cut roughly in half vs. GPT Image 2 | Slower, longer generation times |
| Best for | Social content, rapid prototyping, high-volume batches | Campaign creative, polished product shots, detail-sensitive finals |
| Cost | Same token pricing as Sunburst | Same token pricing as Flare |
The counterintuitive part: Sunburst does not cost more per token. You pay for the extra precision with time, not money. Both models list the same rates on OpenAI's pricing page: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, unchanged from GPT Image 2.
A practical workflow falls out of that split. Prototype with Flare until the composition is right, then run the final one or two images through Sunburst. Speed where iteration matters, precision where it counts.
What is new vs. GPT Image 2
If you are coming from GPT Image 2, the honest summary is "better everywhere, cheaper nowhere":
- Latency is the headline improvement. OpenAI reports up to 50% faster generation with Flare.
- Reference handling and edit precision improved noticeably, which is where most real-world frustration with the previous model lived.
- Token pricing is identical, so per-image costs stay in the same range. A 1024×1024 image runs from under a cent at low quality to roughly five cents at high, depending on settings.
- Quality settings now span low, medium, high, and beyond, giving finer control over the cost and detail tradeoff.
In ChatGPT, the release also added a Sketch tool (draw a rough guide directly in the prompt), templates for formats like posters and merch, inline comments for requesting edits on specific parts of an image, and shareable prompts others can remix with their own photos. Those conveniences live in ChatGPT; they are not automatic API features.
ChatGPT, API, or a browser studio?
Three ways to touch this model family, depending on what you are doing:
- ChatGPT needs zero setup and includes the new Sketch and templates features. Fine for casual use.
- The OpenAI API gives full control over model choice, sizes, quality levels, and edit endpoints, in exchange for a developer account, an API key, and code that handles billing by token.
- A browser-based studio skips the API key and the code. Someone else runs the API and bills you simple credits.
If the third option sounds appealing, we built one. It is a sibling project to this site, run by the same team.
GPT Image 2.5 Studio: try Flare and Sunburst without an API key
GPT Image 2.5 Studio is an independent generator and editor that puts both new models behind one interface. It is not OpenAI's ChatGPT app and not an official OpenAI product; it is a standalone studio that runs the GPT Image 2.5 models and charges simple credits instead of tokens.
What you can do in it:
- Generate from text or edit from references. Bring up to 16 reference images into an edit and specify what should change and what must stay identical.
- Switch between Flare and Sunburst per generation, so the fast and precise workflow above is a dropdown, not a code change.
- Control the output format properly. Aspect ratios from square out to 3:1 panoramas, long-edge sizes up to 3840 px, PNG, JPEG, or WebP, transparent backgrounds when you need them, and up to 4 images per run.
- See the credit cost before you generate, so there is no invoice archaeology afterward.
- Keep a history and feed any result back in as the reference for the next edit.
- No watermark, and every plan allows commercial use.
New accounts start with free credits, and paid plans are monthly subscriptions or one-time credit packs. If you have used the generator on our homepage, which currently runs GPT Image 2, the Studio is the same idea pointed at the new model family.
Prompt tips that work with GPT Image 2.5
The model got better at following instructions; it did not get better at reading minds. Two habits cover most cases.
For generation, name five things: subject, setting, light, composition, and any exact wording. "A ceramic espresso cup on a scarred wooden workbench, morning side-light, shot straight-on at eye level, with the word ROSTER on a paper tag" beats a paragraph of atmosphere.
For edits, name both sides of the change: what should change, and what must be preserved. "Replace the background with a concrete wall; keep the bottle, label text, and lighting direction unchanged." The precision editing in 2.5 rewards this structure directly.
FAQ
When was GPT Image 2.5 released? September 8, 2026, in ChatGPT and the API at the same time.
Is GPT Image 2.5 free? It is included in ChatGPT across all tiers within its normal usage limits. API usage is billed per token at the same rates as GPT Image 2. Third-party studios like GPT Image 2.5 Studio charge their own credits.
Which should I use, Flare or Sunburst? Use Flare for everyday generation and iteration; use Sunburst when detail matters most and you can wait longer. They cost the same per token.
Does it replace GPT Image 2? For quality-sensitive work, effectively yes: same price, better results. GPT Image 2 remains available, and our homepage generator still runs it.
Is GPT Image 2.5 Studio an OpenAI product? No. It is an independent tool built by the same team as this site, using the GPT Image 2.5 models. OpenAI has not reviewed or endorsed it.
Try it
If you want to compare the two models side by side, GPT Image 2.5 Studio is the fastest path: open a tab, type a prompt, pick Flare or Sunburst. To stay on the current stable model, the GPT Image 2 generator on our homepage is right there, with credits and plans on the pricing page. And if what you actually want is a face analysis tool rather than image generation, we made one of those too; see our AI attractiveness test guide.
