Fast enough to iterate
Two to four times quicker than GPT Image 2 in our run. Try three directions in the time one used to take, then keep the best.
OpenAI's fastest image model.
20% off with GPTIMAGE20 GPTIMAGE20 applied
GPT Image 2.5 Flare is OpenAI's fast image model, released on September 8, 2026 alongside Sunburst. OpenAI calls it the default choice for most work: better images than GPT Image 2 at half the latency.
It ranked second on LMArena's text-to-image and image-edit leaderboards in late September 2026, behind only Sunburst, and spelled every word right in our tests, Japanese included. Use it for drafts, social posts, mockups and quick rounds of edits.
Two to four times quicker than GPT Image 2 in our run. Try three directions in the time one used to take, then keep the best.
Carousel covers, thumbnails and ads with the exact headline you typed, in a layout that already fits the feed.
Put a logo on a tote, a label on a bottle or a screen in an app, with text that stays readable.
Swap the background and keep the bottle, the cap and every word on the label. Say what changes and what stays.
One try per model with Fiddl's settings, nothing retouched. Times are from our run on October 3, 2026.
All three spelled every line correctly.
Three natural, believable candids from the same sentence.
Flare added a kicker and two taglines nobody asked for; the other two used only the requested text.
First tries with GPT Image 2.5 Flare, nothing retouched. Tap one to see its prompt.
On our coffee infographic Flare added a kicker and two taglines that were not in the prompt; Sunburst and GPT Image 2 used only the requested text. When the wording matters, end the prompt with "no extra text", or use Sunburst.
Pick Flare when speed matters and everyday quality meets the brief: drafts, variations, social content, mockups. Switch to Sunburst when a face, a product or a line of copy must come through an edit unchanged, or when Flare's result is close but not quite right. Both cost the same per image on Fiddl.
| Model | Flare | Sunburst | GPT Image 2 |
|---|---|---|---|
| Best for | Fast everyday images and edits | Precise edits and polished campaign work | Lettering and layouts on a budget |
| Speed in our test | 27 s | 40 s | 99 s |
| LMArena text-to-image September 24, 2026 | #2 | #1 | #3 |
| LMArena image editing September 29, 2026 | #2 | #1 | #3 |
| Images with the Studio pack | about 125 | about 125 | about 250 |
| Reference images | up to 16 | up to 16 | up to 10 |
| Released |
Condensed from OpenAI's image prompting guide and our own tests.
Launch offer 20% off any pack with code GPTIMAGE20
Code GPTIMAGE20 applied: 20% off every pack
Already on Fiddl? Open GPT Image 2.5 Flare in the app Use code GPTIMAGE20 on your next pack
OpenAI's fast image model for everyday generation and editing, released on September 8, 2026. It takes text and up to 16 reference images on Fiddl and makes images up to 4K in 11 aspect ratios.
Flare for speed: drafts, social posts, mockups and lots of variations. Sunburst for precision: campaign creative, product edits and faces that must stay the same. In our test Flare took about 27 seconds per image and Sunburst about 40; on Fiddl they cost the same.
OpenAI says Flare is higher quality than GPT Image 2 at half the latency, and Flare ranked above GPT Image 2 on LMArena's leaderboards in late September 2026. OpenAI's own prompting guide is more modest and calls the quality comparable, so think of Flare as GPT Image 2 quality or better, much faster.
About 25 Flare images with the $10 Starter pack and about 125 with the $42.50 Studio pack, from $0.28 per image on the largest pack. 4K images cost a little more. During the launch, code GPTIMAGE20 takes 20% off any pack until November 2, 2026.
There is no free tier for Flare on Fiddl. Packs are one-time payments from $10 with no subscription, and the balance stays in your account.
Yes. Add up to 16 reference images, say what should change and what must stay, and Flare edits from them. For difficult edits where a face or a product must be preserved exactly, Sunburst is the stronger choice.
OpenAI's API supports transparent output, but Fiddl currently returns images with a normal background. Ask for a plain white background, then remove it with Fiddl's background remover.
No. Describe what you need and your creation agent writes the prompt and picks the model. Every example here shows the exact prompt we used, if you want a starting point.