GPT Image 2.5 on VdoBloom: Flare vs Sunburst, Credits and Prompts

GPT Image 2.5 comes in two variants on VdoBloom, Flare and Sunburst, and they cost exactly the same: 3 credits at 1K, 5 at 2K and 8 at 4K, for both generation and editing. Flare is the faster, default variant built for everyday work; Sunburst trades some speed for edit precision and fidelity. Sunburst is not a premium tier, so there is no price reason to default to Flare — pick by the job.

Updated October 2026.

What GPT Image 2.5 is

OpenAI released GPT Image 2.5 on 8 September 2026, and it is exposed on VdoBloom as four separate models: Flare and Sunburst, each with its own text-to-image and image-to-image id. Both variants accept the exact same input schema at the same prices, which is the part worth knowing before you pick one.

Compared with GPT Image 2, the pricing shape did not move — it is still 3, 5 and 8 credits by resolution — but two real constraints did. GPT Image 2’s rule that 1:1 could not be rendered at 4K is gone, and 2.5’s own resolution-limited ratios clamp automatically to 1K rather than failing the job. Editing capacity also rose from 10 reference images to 16.

Flare vs Sunburst: the honest difference

Both are the same model family, launched together, with identical inputs and identical prices, whether you generate or edit. The only real difference is what OpenAI optimised for.

Variant	What it is optimised for	Rating on VdoBloom	Credits
Flare	Speed — everyday generation and quick iteration	4.7	3 / 5 / 8 at 1K / 2K / 4K
Sunburst	Edit precision and fidelity, for work that has to hold up under scrutiny	4.8	3 / 5 / 8 at 1K / 2K / 4K

The way to read that table: Sunburst is not a premium tier, and choosing it does not raise your bill. Reach for Flare when turnaround matters, and switch to Sunburst when the specific job benefits from precision — fine detail, tight layout, or on-image text that has to be exactly right.

What it costs, and what the free credits cover

Cost is shown before every generation, and the ladder is the same for text-to-image and image-to-image.

Resolution	Credits per generation or edit
1K	3
2K	5
4K	8

New accounts start with 10 free credits and no card required, which covers three 1K generations (9 credits) with one left over, or two 2K generations. If you are budgeting a project, VdoBloom’s pricing runs from Lite at $15/month for 300 credits to one-time credit packs from $10.39 for 200 credits that never expire — and paid output is watermark-free with commercial rights. At 3 credits a 1K draft, this is one of the cheaper ways to iterate on an image: a fifty-frame concept sweep at 1K is 150 credits.

Technical limits worth knowing before you generate

Resolution: 1K, 2K or 4K. All three are available on both variants and both input modes.
Aspect ratios: thirteen, including auto — 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 27:16, 16:27, 9:8 and 8:9. Four of them (27:16, 16:27, 9:8 and 8:9) are limited to 1K, and VdoBloom clamps the resolution automatically instead of rejecting your request, so you are never charged for a failed job.
Prompt length: up to 20,000 characters, which is far more room than most prompts need and enough for a tightly specified art direction brief.
Editing references: up to 16 input images per edit, up from 10 on GPT Image 2.
Content filtering: GPT Image 2.5 carries VdoBloom’s STRICT content label, matching OpenAI’s provider-side filtering and the same tier as GPT Image 2 and GPT Image 1.5. Strict means literally described scenes with transparent subjects; it is not the tier to pick for anything ambiguous.

Example prompts

These are written for the things GPT Image 2.5 does well: literal, detailed instructions, exact counts, specified positions, and text that has to appear in the image correctly.

Product still with legible text. “Studio product photo of a matte black ceramic coffee jar on a pale oak shelf, label reads ‘Slow Roast’ in a thin uppercase serif, three smaller jars blurred behind it, soft window light from the left, no props, 4:5.”
Concept sweep at 1K. “Minimal app icon, single line-art umbrella in warm terracotta on a cream square, generous margin, flat vector look, no shadow.” Run it three times at 1K for 9 credits and keep the one that reads at 32 pixels.
Multi-reference composite. Upload your product shot, a fabric swatch and a background plate, then: “Place the bottle on the linen swatch in the foreground, background plate behind, keep the bottle’s label and proportions exactly as in image 1, one soft shadow to the right, no added text.”
On-image text with a tight layout. “A square book cover, off-white board, title reads ‘The Long Field’ in letterpress caps across the upper third, author name ‘A. Rowntree’ small in the lower right, one thin rule between them, nothing else on the cover.” This is the kind of job where Sunburst earns its place.
Ultrawide banner. “A 21:9 panoramic illustration of a ridge line at dusk, three pine trees at the far right, empty negative space across the left two-thirds for a headline, muted teal and amber palette, no figures.”
Sequential sprite sheet. “Six-panel sprite sheet of one cartoon fox, same character in all six, front view, three-quarter, profile, crouch, jump, land — flat shading, consistent scale, thin black outline, transparent background.” Six frames is well inside the 16-reference editing ceiling.

Getting better results

Spell out counts and positions. The model follows literal instructions well, so “three jars, the front one sharp” beats “a few jars”. Detail is cheap here because the prompt budget is generous.
Name the ratio explicitly when you are paying for 2K or 4K. Four ratios clamp to 1K by design.
Draft at 1K, deliver at 4K. Because the pricing ladder is 3/5/8 rather than a flat rate, three 1K explorations cost the same as one 4K render, and only the final pass needs the expense.
Put excludes in positive terms. GPT Image 2.5 does not expose a negative-prompt field on VdoBloom; describe the frame you want with the unwanted elements left out of the description (“plain wall, no signage” rather than a list of things to avoid). If you want an actual exclusion field, the Qwen text-to-image models are the ones on VdoBloom that carry one — Qwen Image 3.0 and Qwen Image 3.0 Pro among them.
Use Sunburst for text and layout, Flare for volume. Same price, different trade-off.

Which model to pick

Inside the GPT Image family, choose 2.5 over GPT Image 2 unless you have a specific reason to stay: the price is unchanged, the ratio restriction is removed and editing capacity is larger. Between the two 2.5 variants, Flare is the sensible default and Sunburst is the escalation for precision work.

Outside the family, the comparison people ask about most is against Nano Banana, Google’s model on VdoBloom. Nano Banana is flat-priced at 6 credits with a narrower ten-ratio set and no resolution control; GPT Image 2.5 scales from 3 to 8 credits across thirteen ratios. Flare is cheaper for 1K and 2K work and gives you resolution choice; Nano Banana is simpler if you would rather not think about resolution at all. For a wider view, see the best AI image models in 2026.

Frequently asked questions

Are Flare and Sunburst different prices?

No. Both cost 3 credits at 1K, 5 at 2K and 8 at 4K on VdoBloom, for generation and editing alike. Sunburst is not a premium tier; it trades latency for edit precision and fidelity.

Which variant should I use by default?

Flare, because it is the faster of the two and is built for everyday generation. Switch to Sunburst when the job depends on fine detail, precise layout or on-image text that has to be exactly right.

Does GPT Image 2.5 have a negative prompt field?

Not on VdoBloom. GPT Image 2.5 takes a single prompt of up to 20,000 characters and no separate exclusion field, so describe unwanted elements by describing the frame positively. The Qwen text-to-image models here — Qwen Image 3.0 and Qwen Image 3.0 Pro included — do expose a negative-prompt field.

How many reference images can I use when editing?

Up to 16 per edit, up from 10 on GPT Image 2. That makes multi-frame sprite sheets and composites that genuinely need many sources practical in one pass.

What aspect ratios are available?

Thirteen including auto: 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 27:16, 16:27, 9:8 and 8:9. The last four are 1K-only, and VdoBloom clamps the resolution automatically rather than failing the request.

Try it with the free credits every new account gets: GPT Image 2.5 Flare for speed or GPT Image 2.5 Sunburst for precision, and read how to write a better AI image prompt before your first 4K render.