GPT Image 2 vs Flare vs Sunburst: API test results

Results and settled costs

On the same X route, GPT Image 2 completed all three scenarios on the first submission. Flare completed the product and poster scenarios on the second submission; its editing scenario failed four times. Sunburst failed all four submissions in each scenario before model generation. Missing outputs are shown explicitly below.

The main comparison used 23 submissions: 5 succeeded, 18 failed, and the settled total was 0.060 credits. Each successful X image cost 0.012 credits; failed submissions in this test settled at zero. This small, retry-dependent sample cannot estimate a long-term success rate.

Identical prompts, bounded retries, first successful outputs

We requested size=1024x1024, quality=high, n=1, output_format=png, response_format=url and async=true for all three public model IDs. The key was pinned to X and the effective route was checked against billing. X needs account access. At test time the account had no Mix quote for Flare or Sunburst, so this comparison does not represent Mix.

We first submitted one request per model and prompt. After investigating terminal failures, the main X comparison allowed at most four total submissions per combination, with 20, 40 and 60 seconds of backoff before the additional attempts. We stopped on the first successful output and did not reroll for appearance. Unknown or still-running tasks were not resubmitted. Earlier diagnostic requests are disclosed separately below.

Times below measure client submission to observed terminal status, including network time and five-second polling intervals. They exclude retry backoff and the separate diagnostic runs. We publish every attempt duration rather than only the successful attempt. Shared route names do not prove identical upstream infrastructure. Requested high quality is not independent proof of an upstream quality tier.

Product photography: the same blue mug prompt

Inspect the handle position, silhouette, ceramic surface and empty space. These are unretouched original outputs; the smaller WebP previews are only resized for page delivery.

A studio product photograph of a single cobalt-blue ceramic mug on a warm ivory background. The handle is on the right. Show the entire mug, centered with generous empty space, soft lighting and a subtle shadow. Square composition. No text, logos, watermark or other objects.

No image returned

These requests did not produce an image. Attempt details record network-health or upstream-capacity failures; visual quality cannot be assessed.

Text poster: English, Chinese and exact line breaks

The requested strings are SLOW MORNINGS, 咖啡与好时光 and SATURDAY 9 AM, each on one line. Check spelling, the three-line requirement and object placement separately. The original prompt is kept unchanged in every language version.

Design a flat square poster for a fictional coffee shop. Warm ivory background, navy typography, one blue ceramic mug in the lower half and one small orange circle in the upper right. Include exactly these three text lines, each on one line, with correct spelling and no other text: "SLOW MORNINGS", "咖啡与好时光", "SATURDAY 9 AM". Keep all letters inside generous margins. Do not show a frame, wall or paper mockup.

No image returned

These requests did not produce an image. Attempt details record network-health or upstream-capacity failures; visual quality cannot be assessed.

Reference editing: preserve the mug, replace the background

Every editing request used one identical reference: the public blue-mug sample from our September 17 test. The instruction changes the background to pale gray and adds one orange sphere on the left. Only GPT Image 2 returned an edited image in the main X cohort; there is no three-model reference-fidelity result yet.

Edit the reference image. Keep the same blue ceramic mug, including its exact shape, blue glaze, proportions and right-side handle. Replace only the background and surface with warm pale-gray seamless studio paper. Add one small matte orange sphere on the left of the mug, clearly separate from it. Keep the whole mug inside the square canvas. No text, logo or watermark.

No image returned

These requests did not produce an image. Attempt details record network-health or upstream-capacity failures; visual quality cannot be assessed.

No image returned

These requests did not produce an image. Attempt details record network-health or upstream-capacity failures; visual quality cannot be assessed.

What the failed requests do—and do not—show

Most Flare and Sunburst failures were network-health rejections before generation. One Sunburst X attempt returned an upstream capacity error; one official Flare diagnostic returned an unavailable-channel error. These show delivery problems during the test, not an absence of model capability. With no Sunburst output, a three-model visual ranking is not possible.

Before the bounded X retries, an official-route diagnostic repeated the product and poster prompts and tried the background edit. GPT Image 2 completed three images; the other six requests failed. Those three official outputs cost 0.098775 credits in total and are listed separately in the evidence file. They are not mixed into the X gallery or used to compare speed.

An earlier transparent-background edit also failed once per model. One GPT Image 2 upstream explicitly rejected background=transparent; Flare and Sunburst were blocked before generation. These three failures cost zero. Including these diagnostics, the complete session used 35 submissions, produced 8 images and settled 0.158775 credits. A successful task is checked separately from its visual quality and file format.

Cost and integration questions

Does the test prove Flare is faster? No. The observed successful Flare requests took about 34 seconds, but followed failures and used different infrastructure. A speed ranking needs repeated controlled runs, queue measurements and a disclosed sampling method.

Is 0.012 credits the price for every user? No. It was the settled X price for this account and date, with 1 credit equal to US$1. Check live account pricing and permissions before running. The GPT Image 2 price page separately explains catalogue per-image budgets and token-based generation versus editing costs.

Should an HTTP 202 response trigger a retry? No. Save the task ID and poll until a terminal state. Only decide whether to submit again after confirming the previous task failed; a second successful generation is a second billable image.

Model-specific parameters

Public configuration snapshot; not live availability. Follow each model page for its parameters, supported routes and reference limits.

• GPT Image 2 openai/gpt-image-2
• GPT Image 2.5 Flare openai/gpt-image-2.5-flare
• GPT Image 2.5 Sunburst openai/gpt-image-2.5-sunburst

Related guides

• Understand routes, prices and available balance
• Send reference images without avoidable input errors
• Diagnose image API errors before resubmitting
• GPT Image 2 vs Nano Banana 2 vs Pro: six real outputs

References

• API reference