GPT Image 2.5 on Genfire: Sunburst vs Flare, and When to Use Each

Two Tiers, One Price

OpenAI did not ship GPT Image 2.5 as a single model. It shipped two variants with identical controls and identical pricing that differ only in where they spend their time — and both are live on Genfire.

• GPT Image 2.5 Flare is OpenAI's recommended default: higher quality than GPT Image 2 at roughly 50% lower latency. It is the tier for everyday, social, product, and high-volume work.
• GPT Image 2.5 Sunburst is the precision tier: extra fidelity on intricate detail and tighter control across edits, at longer generation times.

On Genfire both tiers are priced the same as GPT Image 2, which makes the choice purely about the job rather than the budget. There is no financial reason to stay on the older model: Flare is a straight upgrade in both quality and speed, and Sunburst is an option GPT Image 2 never offered.

The full spec sheet is at /gpt-image-2-5. This post is about which tier to pick, and where the output goes afterwards.

What Both Tiers Do

The two variants share a schema, so anything you learn on one transfers to the other:

• Text-to-image and edit. Supplying a source image routes the request to the edit endpoint automatically.
• Up to 16 reference images per request, so a product, a model, a background, a logo, and a type reference can all inform one generation instead of being stitched together across several passes. (The REST endpoint's image_urls array caps a multi-image edit at 14.)
• Masked inpainting. Paint a region, change only that region, leave the rest of the frame untouched — the way you fix one element of an otherwise finished image without regenerating it.
• Aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4.
• Up to 4 images per request.
• Quality: low, medium, high. Quality is billed on the same multiplier table GPT Image 2 uses, so a low round is cheap and only the approved direction gets rendered at high. Draft low, ship high.

Flare: The One You Should Default To

Flare is the tier that changes how you work, because latency is what decides whether you generate one candidate or four.

At roughly half GPT Image 2's generation time, a four-up comparison stops being something you set up and walk away from. You write the brief, get four reads on it, kill three, and refine the survivor — in the time a single GPT Image 2 render used to take. That loop is the whole point of a fast model, and it is worth more on most briefs than a marginal gain in fine detail.

Reach for Flare when you are:

• Filling a social calendar. On-brand graphics across 1:1, 9:16, and 16:9 from one concept, at volume.
• Iterating on ad variations. Five headlines, three backgrounds, two product angles — the combinatorics are the job, and speed is what makes them affordable in time.
• Exploring a direction. Early rounds where you do not yet know what the image is.
• Doing routine product shots. Clean lighting, accurate materials, a background that stays out of the way.

Sunburst: For Briefs That Are Unforgiving

Sunburst takes longer and spends that time on two things: intricate detail, and control across edits. Both matter in a narrow set of jobs — but in those jobs they matter a lot.

Reach for Sunburst when:

• The layout is text-heavy. Packaging with forty words of copy that all have to be right, a pharmacy shelf where every label on every box has to read cleanly, an infographic whose callouts have to land in the right places.
• The edit has to be surgical. Swap the jacket, keep the pose, the hair, and the studio light exactly as they were. Tighter edit control is precisely what "change this one thing and disturb nothing else" needs.
• The image is the final deliverable. Campaign creative and hero product imagery that will be looked at closely, printed, or blown up.
• Fine structure is the subject. Jewellery, watch faces, textiles, engraving, circuitry — anything where the detail is the picture.

The sensible habit is to draft on Flare and switch to Sunburst for the final pass. On Genfire the two sit one chip apart in the same model picker, so switching costs a click and the prompt carries over unchanged.

How It Compares

Nano Banana Pro stays in the mix for a reason: it is the premium photoreal pick from Google's side of the catalog, and it reads prompts differently enough that a brief that stalls on one family often lands on the other. It has no quality dial, though, so there is no cheap draft round on it — which is exactly what Flare's low tier gives you.

Where the Output Goes

An image generated on Genfire is a normal asset the moment it exists: it lands in your library, and every downstream tool takes it from there.

• Ad creative. A Flare batch is a set of hooks; a Sunburst pass is the one you actually run. Both feed straight into video as a start frame or a reference, which is the fastest path from a still to a moving ad — see /ai-ad-generator.
• Picture books. Genfire's book pipeline renders its pages on GPT Image 2, which sits on the same price table as 2.5. Everything you work out about prompting one — how to hold a character's face across spreads, how to keep the type legible — carries directly across.
• Influencer stills. Influencer Studio's photo generation runs on GPT Image 2 as well, so a look you develop in the image studio transfers into a model's feed without re-learning the prompt style.
• Everything else. Upscaling, background removal, the workflow canvas, and the Layers editor all accept the output, and the same brief can be A/B tested against Nano Banana, Seedream 4.5, or Muse Image on one credit balance.

How to Run It

In the browser studio

Open the AI image generator and pick GPT Image 2.5 Flare or GPT Image 2.5 Sunburst from the model list. Set the aspect ratio, the count (up to four), and the quality tier, then generate. Drop in a source image and the studio switches to the edit endpoint; paint a mask to constrain the change to one region. Both tiers are also selectable in the workflow canvas, the Layers editor, and the agent's generation settings.

Through the REST API

Two model ids: image.gpt_image_25_flare and image.gpt_image_25_sunburst. Passing image_url or image_urls routes automatically to the edit endpoint; adding mask_url makes it a masked inpaint.

curl -X POST https://api.genfire.ai/v1/images/generations \
  -H "Authorization: Bearer $GENFIRE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "image.gpt_image_25_sunburst",
    "prompt": "A pharmacy shelf set for a fictional brand \"Alder & Vale\" — six product boxes, one consistent brand system, every label readable.",
    "aspect_ratio": "3:4",
    "quality": "high",
    "count": 2
  }'

Keys, the full parameter reference, and cost estimation live at /developers.

Through the MCP server

Point Claude, ChatGPT, Cursor, or any MCP client at mcp.genfire.ai and call genfire_generate_image with model: "image.gpt_image_25_flare" (or _sunburst). The tool takes the same prompt, aspect_ratio, count, quality, image_url / image_urls, and mask_url fields as the API.

From the CLI

npm i -g @genfire/cli && genfire auth login

genfire generate image "Photoreal flat-lay for a coffee subscription — exactly four bags, one scoop, one printed card reading 'ROAST DATE: TODAY'" \
  -m image.gpt_image_25_flare -a 4:3 -n 4 --quality medium

Repeat -i to attach source images for a multi-image edit; local paths upload automatically.

Pricing

Genfire is pay-as-you-go. Credit packs start at $19 for 1,000 credits, purchased credits last 12 months, and every purchase includes a commercial license; optional monthly plans start at $29 a month. Both GPT Image 2.5 tiers are billed exactly as GPT Image 2, with quality applying the same multiplier on top — so the cheapest way to work is a low-quality exploration round followed by one high-quality render of the direction you picked. The studio and POST /v1/models/estimate-cost quote the exact number before anything runs, and current rates are on the pricing page.

The Short Version

Default to Flare. It is better than GPT Image 2 and roughly twice as fast, at the same price, which makes it the correct choice for almost every brief. Switch to Sunburst when the image has a lot of small text on it, when an edit has to move one thing and nothing else, or when the file is the deliverable rather than a step toward one.

Both are in the image studio now, the spec sheet is at /gpt-image-2-5, and the model they replace is at /gpt-image-2.