GPT Image 2.5: Flare vs. Sunburst — what's different and what OpenAI's new model can do

What is GPT Image 2.5

GPT Image 2.5 is OpenAI’s set of image models released on September 8, 2026. In ChatGPT they power ChatGPT Images 2.5 — on every plan, on the web, in the apps and in Codex. In the API they’re two separate models:

• Flare (gpt-image-2.5-flare) — a compact model optimized for speed, the default choice;
• Sunburst (gpt-image-2.5-sunburst) — the base model optimized for quality, for detailed work and edits.

The previous version, GPT Image 2, is still in the API. For the whole family, see the GPT Image overview.

What GPT Image 2.5 can do

Sharp detail and natural light. OpenAI lists the key improvements: sharpness, lighting, textures and a style that doesn’t drift from frame to frame.

Reference fidelity. The model holds recognizable objects and people from uploaded photos more reliably — useful for series built around one character or product.

Multi-step edits. Each new step damages less of what already works. Sunburst is more precise here than Flare.

Six quality levels. low, medium, high, xhigh, max and auto. xhigh and max are exclusive to 2.5 — OpenAI suggests turning them on when high doesn’t meet the bar.

Custom sizes. The recommended sizes are 1024 × 1024, 1536 × 1024 and 1024 × 1536, but any multiple of 16 px up to 3840 px per side works. OpenAI considers anything above 2560 × 1440 experimental.

Transparent backgrounds. Officially supported in PNG and WebP, unlike GPT Image 2, where the feature is still in preview.

Mask and multi-image editing. The edit endpoint accepts several images, and the mask applies to the first one. The mask needs an alpha channel and must match the source in size and format.

Partial previews. With streaming, the model returns up to 3 draft frames before the final one — each costs an extra 100 tokens.

Sketches and comments in ChatGPT. ChatGPT’s interface added sketching as a prompt, templates and editing through a comment pinned to a specific part of the image.

Flare vs. Sunburst

	Flare	Sunburst
Type	compact, optimized for speed	base, optimized for quality
Quality	comparable to GPT Image 2	higher
Speed	up to 50% lower latency than GPT Image 2	slower than Flare
Multi-step edits	more reliable than GPT Image 2	the most precise
Best for	everyday generation, bulk jobs	ad campaigns, product shots, complex edits
Price per token	$30 per million output	$30 per million output

OpenAI’s announcement says Flare beats GPT Image 2 on quality, while its prompting guide calls the two comparable. OpenAI customer Manus rated Flare 2–4 times faster than GPT Image 2. Equal per-token pricing doesn’t mean an equal price per image: the models spend a different number of tokens on each one.

OpenAI suggests choosing like this: for a demanding task, start with Sunburst and confirm it delivers. Then run the same prompts through Flare and switch only if the quality holds and latency drops.

What GPT Image 2.5 is good for

• Ad creative and product shots with several rounds of edits — Sunburst.
• Posters, covers and infographics with lots of text.
• Products and stickers on transparent backgrounds for marketplaces and design.
• Series with one character or product built from references.
• Bulk generation — Flare is faster and suits hundreds of variations.

GPT Image 2.5 vs. alternatives

	GPT Image 2.5	GPT Image 2	Nano Banana Pro	Seedream 5.0
Developer	OpenAI	OpenAI	Google DeepMind	ByteDance Seed
Released	Sep 8, 2026	Apr 21, 2026	Nov 20, 2025	Pro Jul 8, 2026
Max size	3840 px	3840 px	4K	Lite — 4K, Pro — 2K
Quality levels	6, up to max	4, up to high	1K, 2K, 4K	1K–4K depending on version
Transparent background	yes	preview	no	Pro — layers with transparency
Editing	mask, comments	mask	by prompt	point, box, lasso, hex
In AI2Face	yes	yes	yes	yes

GPT Image 2.5 is the most convenient model for text, UIs and PNGs without a background. Nano Banana Pro is stronger with complex people scenes and likeness across several photos, and Seedream 5.0 Pro with layered layouts.

How to write prompts for GPT Image 2.5

1. Structure: scene → subject → details → constraints. Split long prompts into labeled sections.
2. Purpose and format: “a product photo for a listing, 1:1,” “a vertical poster.”
3. Visible details: materials, lighting, color, medium. If you want photorealism, say so.
4. Quote text verbatim with its position and typeface, plus “no other text.” For small text, use medium or higher.
5. Be exact about framing: “full body, feet visible.”
6. In edits: “change only X,” then list what must stay and what’s off-limits — logos, watermarks.
7. Don’t rely on the prompt for pixel-perfect areas: if a region must stay identical, composite the approved edit into the original.

Compare Flare and Sunburst at the same quality level. If the result falls short, raise the quality; if it passes, try lower settings to save time.

How to use GPT Image 2.5 in AI2Face

The AI2Face generator has both versions — GPT Image 2.5 Flare and GPT Image 2.5 Sunburst — alongside Nano Banana Pro and Seedream 5.0.

1. Sign up — new accounts get starter credits.
2. Open the generator and pick Flare or Sunburst.
3. Write a prompt, and upload references or an image to edit if needed.
4. The credit cost is shown before you start. Credits don’t expire.

Every image on this page was made in AI2Face with GPT Image 2.5 Sunburst; the comparison diptych also used Flare.

Limitations

• Large images are experimental. OpenAI doesn’t officially guarantee anything above 2560 × 1440.
• Masks are guidance. The model can spill past the mask edges.
• Sunburst speed. Complex prompts can take up to 2 minutes.
• Moderation. Requests and results are filtered under OpenAI’s policies.
• Mixed messaging on Flare. The announcement says it beats GPT Image 2 on quality, the guide says it’s comparable. Test on your own tasks.

Bottom line

GPT Image 2.5 is OpenAI’s best image model: Flare handles everyday work quickly and cheaply, while Sunburst covers ads, product shots and multi-step edits. Transparent backgrounds and the xhigh and max levels make it better than GPT Image 2 in almost every way. For scenes with people and likeness across several photos, compare the result with Nano Banana Pro; for layered layouts, with Seedream 5.0 — all in the same AI2Face generator.