GPT Image 2.5 Operator Guide: Flare vs. Sunburst, 32k Prompts, and Native Transparent AI Workflows

For a long time, generative image models gave marketing teams the same compromise: striking pictures that drifted away from the brief, or hours spent masking, fixing edges, and stitching assets together in other tools.

OpenAI's GPT Image 2.5 (the API snapshot is dated September 8, 2026) is aimed at that gap. It ships as two models, Flare for everyday generation and Sunburst for precise edits, and fal's published API schema for them lists transparent backgrounds, very long prompts, and edits that use reference images. Both are available in MITPO Creative Studio on paid plans.

This guide separates what OpenAI and fal document from what is our own experience, and shows how to brief the model for commercial creative.

Key Takeaways (AEO Direct Summary)

• Two tiers: OpenAI describes Flare as fast, high-quality everyday image generation and Sunburst as the model built for editing precision. In MITPO they appear as "GPT Image 2.5" and "GPT Image 2.5 Precise."
• Transparent backgrounds: the model's API accepts a background setting of transparent, which a design pipeline needs for cutouts, stickers, and product shots on changing backgrounds.
• Long prompts: fal's published schema for the model accepts prompts of up to 32,000 characters, enough for a full creative brief.
• Reference editing: the edit endpoint takes reference images and an optional mask. Creative Studio attaches up to four reference images.

Flare vs. Sunburst: Choosing the Right Model

Both are listed at the same token rates on OpenAI's model pages, so the choice is about the job, not the price. In MITPO, 4K output costs more than standard size, and the model picker always shows the live cost.

When to use Flare

Use Flare when you are exploring. If the team is brainstorming a dozen angles for an ad sprint or drafting visuals for a newsletter, the fast model lets you test directions cheaply and keep the best two.

When to use Sunburst

Use Sunburst once the direction is chosen and you are refining. Its documented strength is that an edit changes what you asked for and leaves the rest alone, which is what a product shot or a branded layout needs.

Transparent backgrounds: what to know before you plan around them

Cutouts have always meant a second tool: generate on a backdrop, then mask it. GPT Image 2.5 can skip that step at the model level. fal's schema for the model lists a background setting with the values auto, transparent, and opaque, and the transparent option is what produces cutouts you can drop onto a landing page, a story overlay, or a mockup.

A caveat that matters for planning: transparency only happens when the app that calls the model switches that setting on. Creative Studio does not expose a transparent-background switch today, so test a sample asset before you build a workflow that depends on true transparent output.

The 32,000-character prompt: a brief, not a keyword list

Longer prompts let you write the brief the way you would for a photographer: brand style, composition, subject details, and output requirements in one place. fal's schema for the model sets the prompt limit at 32,000 characters.

A longer prompt is not automatically a better one. Specific beats long: name the lighting, the lens, the exact text to print, and what must not change. Here is a structure to adapt:

[BRAND STYLE]
Brand: Lumina Botanical Skincare (an example brand)
Palette: forest moss green, warm oat, warm gold
Tone: editorial, organic, minimal

[COMPOSITION AND CAMERA]
Eye-level product packshot, 85mm lens, soft directional daylight from the left.
Surface: matte stone pedestal.

[SUBJECT]
Product: 50ml amber glass dropper bottle with a matte paper label.
Label text, exactly: "LUMINA" / "Botanical Repair Serum" / "50ml"

[OUTPUT]
Plain light background. Keep a soft ground shadow.

Always check printed text letter by letter. Even a strong model can slip on small type, so review labels before anything ships.

Reference images and edits

Subject drift is the classic problem of scaling AI creative: variant B changes the pack shape, variant C changes the mascot's face. The model's edit endpoint takes reference images and an optional mask so you can say what to keep and what to change. The API accepts up to 16 reference images per request on fal. In MITPO you can attach up to four, which covers the usual set:

• the product photo, to hold packaging shape and logo placement
• a brand color card
• a lighting or environment reference
• a layout sketch for subject placement and negative space

Pick Sunburst for the edit pass, describe the change, and state what must stay the same.

4 Prompt Templates to Adapt

1. E-commerce beauty packshot (Sunburst)

Commercial macro studio photograph of an organic facial mist bottle on a polished river stone pedestal. Soft studio rim lighting along the glass curve, fine moisture droplets on the bottle. The label reads "AURA ORGANICS" in a modern geometric sans-serif. Sharp product clarity, 85mm lens look.

2. Creator-style product ad (Flare)

An authentic phone photo with direct on-camera flash. A creator holds an unlabelled skincare jar toward the lens, arm extended, imperfect framing, a lived-in bedroom softly behind. Cool flash white balance, hard flash shadow, visible phone-camera grain.

3. SaaS dashboard element (Flare)

Isometric floating analytics card with a rising trendline and circular metric gauges, soft matte pastel accents, clean volumetric lighting. Rendered as an isolated graphic for a landing page hero.

4. Food and beverage lifestyle ad (Sunburst)

High-speed action photograph of an iced matcha latte in a ribbed glass tumbler. Swirling green and cream layers, clear ice cubes catching afternoon sunlight, fresh mint leaves suspended mid-splash. Crisp motion freeze, shallow depth of field.

Treat these as starting points. Run each once, read the output critically, and adjust one variable at a time.

What it costs

OpenAI lists GPT Image 2.5 at $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. In MITPO you pay in credits instead: 6 credits per image at standard size on both models, more at 4K. The free plan includes 100 credits at signup, once, to keep. See Pricing for current plans.

Start with GPT Image 2.5

Both GPT Image 2.5 (Flare) and GPT Image 2.5 Precise (Sunburst) are available on paid plans in Creative Studio:

• Test prompts and compare outputs in the interactive Creative Studio.
• Read the workflow guide in the Creative Studio docs.
• Check credit allocations on the MITPO Pricing page.

FAQ

What is the difference between GPT Image 2.5 Flare and Sunburst?

OpenAI positions Flare as fast, high-quality everyday image generation and Sunburst as the model for editing precision. In MITPO they appear as GPT Image 2.5 and GPT Image 2.5 Precise.

Does GPT Image 2.5 support transparent backgrounds?

Yes at the model level: fal's schema lists transparent as a background option. Whether an app exposes it is separate, and Creative Studio does not have a transparent-background switch today.

How long can a GPT Image 2.5 prompt be?

fal's schema for the model allows prompts of up to 32,000 characters. Specific instructions still beat long ones, so write a structured brief rather than padding.

How many reference images can I use?

The model's edit endpoint accepts up to 16 reference images on fal. Creative Studio lets you attach up to four.

What does GPT Image 2.5 cost?

OpenAI lists $5, $8, and $30 per million text-input, image-input, and image-output tokens. In MITPO it is 6 credits per image at standard size, with 4K costing more.