GPT Image 2.5 Flare vs Sunburst: Token Cost and API Choice

Flare and Sunburst share the same published token rates on PiAPI. Choose the operation you need—generation or editing—then select a variant and track actual usage. A shared rate card does not guarantee the same final bill for two different requests.

What differs between Flare and Sunburst?

PiAPI describes Flare as its faster everyday option and Sunburst as its precision-oriented option for generation and editing. These are descriptions from the public model guide, not results from an independent comparison in this article.

Keep the explicit model suffix in API requests. The separate gpt-image-2.5 route without a suffix has different control semantics; it is not shorthand for selecting Flare or Sunburst.

The same three token rates apply to both variants

These published rates were checked on September 22, 2026. Neither variant has a price advantage at identical usage in these three categories. Your total still depends on the tokens used, so do not convert the rate card into a fixed price per image.

Worked example: assumed usage, not a final bill

For illustration, assume 2,000 text-input tokens, 1,000 image-input tokens, and 4,000 image-output tokens. Divide each count by one million before applying the corresponding rate:

Text input:   2,000 / 1,000,000 × $3.75  = $0.0075
Image input:  1,000 / 1,000,000 × $6.00  = $0.0060
Image output: 4,000 / 1,000,000 × $22.50 = $0.0900
Illustrative total:                     $0.1035

The sum is $0.1035 in this illustrative calculation. It is not a quoted task price or the final bill. Use actual usage and the current account billing record; do not guess the text/image split from a combined input-token count.

Generation and editing use separate endpoints

Both variants support POST /api/v1/images/generations/async and POST /api/v1/images/edits/async on https://api.piapi.ai. The async header is X-API-KEY. Choose generation for a prompt-only input and editing when a reference image is part of the request.

The asynchronous API documentation describes edits using a multipart image file or a JSON image string containing an inline data URL. Use that documented wire format when integrating directly. Keep the source image and edit receipt separate from any preceding generation.

Persist the task ID, poll GET /api/v1/task/{task_id} with a bounded client wait, and handle failed as terminal. The async documentation shows completed base64 images under output.data[0].b64_json. Resume the saved task after a polling timeout instead of submitting the same request again.

For executable submission, polling, image decoding, and edit examples, use the existing GPT Image 2.5 first generation and edit walkthrough. This comparison focuses on variant selection and token budgeting.

Choose a variant for your workload

1. Decide whether you need a new image or an edit to an existing image.
2. Select Flare or Sunburst using the documented positioning as a starting point. Evaluate representative outputs yourself when your application needs a measured quality or speed comparison.
3. Use a documented quality value: auto, low, medium, or high. Keep size, quality, and input images consistent when comparing variants.
4. Record the exact variant, operation, task ID, and available usage breakdown. Compare recorded costs instead of assuming that matching rates imply matching token use.

The GPT Image 2.5 model page and API documentation provide the current controls and contract. No independent latency or quality benchmark is presented here.