GPT Image 2.5 ships as two sibling models on Callirra. They share the same prompt limit (20,000 characters), reference capacity (up to 16 images) and price matrix, but target different looks. Choose a resolution and quality for each run; the price appears before you generate.
Updated September 27, 2026: GPT Image 2.5 now supports selecting resolution and quality together. Prices below reflect the current 15-combination matrix.
Flare — photographic
Flare is the realistic one: natural skin and material texture, believable lighting falloff, product-shot framing. Reach for it when the image is supposed to look like it was captured rather than composed.
Sunburst — stylised
Sunburst leans into illustration, poster and graphic treatments. It is the better starting point for thumbnails, covers and anything where colour blocking matters more than photorealism.
Two controls, one price
- At 1K, low / medium / high / xhigh / max cost 3 / 4 / 6 / 9 / 13 credits.
- At 2K, the same quality tiers cost 5 / 7 / 10 / 15 / 22 credits.
- At 4K, they cost 8 / 11 / 16 / 24 / 35 credits.
Resolution and quality work together. A 2K + high image costs 10 credits; 4K + max costs 35. Flare and Sunburst use the same matrix, and the final cost is shown next to Generate.
A 1K + low draft costs 3 credits. Iterate on composition there, then keep the prompt and raise resolution or quality for the image you plan to use.
A workflow that keeps the bill small
- Explore at 1K + low (3 credits) until the composition is right — a 15-credit weekly allowance covers five attempts.
- Lock the prompt, then move to 2K + high (10 credits) for delivery; 4K + high (16 credits) only when you need the pixels.
- Use image-to-image with a reference instead of re-describing a look in text; it converges in fewer attempts.
Does the style change the price? No — the tier does
- Same settings, same price: a 2K + high Flare frame and a 2K + high Sunburst frame are both 10 credits. Switching style carries no surcharge.
- Resolution and quality jointly set the bill: from 3 credits for 1K + low to 35 for 4K + max. You can increase either control independently.
- For a large, high-detail frame, select the resolution and quality you need together. The generator shows the exact cost before submission.
- Reference images do not change the price. Up to 16 references are accepted at the same per-image cost, so image-to-image with several locked references is cheaper than re-rolling text prompts until one lands.
What our own runs show — and what they do not
We publish a 50-frame GPT Image 2.5 atlas: each historical frame has its prompt and recorded settings, including the output tier used at the time. It is a visual comparison of the current 15 price combinations. Flare tends toward photographed skin, metal and glass; Sunburst leans into graphic colour and illustration.
What the atlas does not show is a quality ranking. We are not going to claim one tier is "better" from our own small sample — the honest statement is that Flare reads as photographed and Sunburst reads as illustrated, and which of those you want is a decision about the picture, not about the model.
A five-minute test that settles it for your use case
- Take one prompt you actually care about, with specific objects, materials and any in-image text.
- Render it at 1K + low on both Sunburst and Flare — 3 credits each, so the experiment costs 6.
- Look at three things: skin and material texture, how the light falls off, and whether the text is legible.
- Whichever wins, keep that prompt and re-render the winner at 2K or 4K for delivery. The prompt is portable between the two — nothing needs rewriting.
Every card in the published atlas loads its prompt straight into the generator, so you can reproduce any of the 50 frames rather than take our summary of them on trust.