Castaly · Blog · Models
GPT Image 2.5 Flare, Sunburst and GPT Image 2 Compared
Flare and Sunburst share one price; GPT Image 2 costs about 29% less per token. Estimates by size and quality, what real runs cost, and which one to pick.
Castaly has three GPT Image models. Two of them, GPT Image 2.5 Flare and Sunburst, share one price and one set of controls, so choosing between them is not about money. The third, GPT Image 2, is cheaper per token and has fewer options.
At a glance
gpt-image-2.5-flare |
gpt-image-2.5-sunburst |
gpt-image-2 |
|
|---|---|---|---|
| Documented purpose | Everyday generation and editing | Higher-fidelity, higher-control work | — |
| Output tokens, per million | $31.98 | $31.98 | $22.62 |
| Input tokens, per million | $5.33 | $5.33 | $3.77 |
| Quality in the Studio | Low, medium, high | Low, medium, high | Auto, low, medium, high |
| Default quality | Medium | Medium | Auto |
| Transparent background | Yes | Yes | No |
| PNG, JPEG or WebP, with compression | Yes | Yes | No |
| Moderation setting | Yes | Yes | No |
| Custom sizes | Yes | Yes | Yes |
| Reference images | Up to 16 | Up to 16 | Up to 16 |
| Images per request | Up to 10 | Up to 10 | Up to 10 |
Flare and Sunburst's descriptions come from their docs pages. Those pages also document xhigh and max quality and a mask for edits through the API.
What the Studio estimates
The estimate before you generate depends on size and quality:
| Model | Size | Low | Medium | High |
|---|---|---|---|---|
| Flare / Sunburst | 1024 × 1024 | $0.006 | $0.056 | $0.225 |
| Flare / Sunburst | 1536 × 1024 | $0.005 | $0.044 | $0.176 |
| GPT Image 2 | 1024 × 1024 | $0.004 | $0.040 | $0.159 |
| GPT Image 2 | 1536 × 1024 | $0.004 | $0.031 | $0.124 |
Two things stand out. Quality is the big lever, with high costing four times medium and medium about nine times low. And 1536 × 1024 is estimated below the square size, so a landscape or portrait image does not cost more than a square one.
On GPT Image 2, auto lets the provider choose the quality, so its estimate assumes high.
What real runs cost
The estimate is a hold. The final charge uses the tokens reported for the finished image, and in our recorded runs at 1536 × 1024 it came in well below the estimate:
| Run | Quality | Output tokens | Input tokens | At today's price |
|---|---|---|---|---|
| Flare, fox illustration | Medium | 343 | 47 | about $0.011 |
| Flare, whale collage | Medium | 343 | 46 | about $0.011 |
| Flare, hands and mug | Not set | 158 | 56 | about $0.005 |
| Flare, paper wave | Not set | 601 | 99 | about $0.020 |
| Sunburst, film ribbon | Not set | 1,372 | 136 | about $0.045 |
| GPT Image 2, marble arch | Not set | 158 | 124 | about $0.004 |
These are single runs, not an average, and different prompts and models produced very different token counts. All three medium-quality Flare runs reported exactly 343 output tokens, so at a fixed quality Flare looks predictable. With quality not set, the same model reported 158 and 601.

Flare: the default
Start here. Flare covers photoreal work and illustration, and it has the full set of output controls. Both of these came from short, plain prompts:
Flat vector illustration of the pour-over coffee process arranged as four simple icons on a cream background: grinder, kettle, dripper, cup. Two-tone palette of terracotta and charcoal, clean geometry, no text or labels.

Sunburst: for a shot that matters
Sunburst costs the same per token and is documented as the higher-fidelity, higher-control option. Our one recorded run, with quality not set, used four times the output tokens of the medium-quality Flare runs at this size, so it can settle higher even though the rate is identical.
The open top deck of a car park at dusk, wet asphalt still throwing back the sky. A woman in a cobalt-blue structured trench coat has just thrown a stack of contact sheets and film strips out of her hands. The strips are not falling: they have started to unwind and climb around her in a slow rotating ribbon, every frame lit a different neon colour …

GPT Image 2: lowest rate, fewer options
GPT Image 2 is about 29% cheaper per token. It has no transparent background and no output format or compression controls in the Studio. It suits high-volume work where the image goes on as a plain PNG.

Which one to use
- Most images: Flare at medium quality.
- A cut-out for a design: Flare or Sunburst with a transparent background. No other model in Castaly offers it.
- A hero image worth extra attention: Run the prompt on both Flare and Sunburst and compare.
- Large batches with no special output needs: GPT Image 2.
- Rough exploration: Any of them at low quality, estimated under a cent, then raise the quality on the composition you keep.
To compare them with Seedream, Wan and Gemini, see all ten image models.
FAQ
What is the difference between Flare and Sunburst?
They share a price and a set of Studio controls. The docs describe Flare as the everyday default and Sunburst as the higher-fidelity, higher-control option. In our one Sunburst run, the output token count was higher, so it settled higher.
Is GPT Image 2 still worth using?
Yes, if you want the lowest per-token rate in the family and do not need transparency or output format controls.
Why was I charged less than the estimate?
The estimate reserves the full token budget for that size and quality. The final charge uses the tokens the provider reports, and the difference goes back to your balance.
Which GPT Image model supports a transparent background?
Flare and Sunburst. GPT Image 2 does not offer it in the Studio.