Is GPT Image 2.5 really 50% faster? Time your own calls on Sume

No single number answers this. OpenAI's announcement dated 2026-09-08 describes Flare as 50% faster than Sunburst, prioritising speed for everyday prompts, while Sunburst takes longer for precision. That compares two models. Your prompt, quality tier and size decide the real wait, so time your own calls: a 20-line script gives you a median in a few minutes.

What the sources say, and don't

The announcement states the Flare versus Sunburst comparison and says nothing you can budget as seconds. OpenAI's image generation guide adds that complex prompts may take up to 2 minutes. Neither source gives a median latency for a given size and quality, so any table of seconds you find elsewhere is a measurement of somebody else's prompts.

What you can say safely is directional: Flare for speed, Sunburst for precision, low quality before max, smaller before larger. That is also why the model choice and the quality tier belong in config.

A small timer

The script posts the same prompt to both models with mode: "async" omitted, so the sync path is measured, and records wall-clock time. A 202 counts as slower than 30 seconds and is noted. Run it ten times per setting and look at the median, not the best run.

import os
import statistics
import time
import requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
PROMPT = "A studio photo of a ceramic mug on a wooden table"

def timed(model: str, quality: str) -> float:
    t = time.perf_counter()
    r = requests.post(
        "https://api.sume.com/v1/images",
        headers=H,
        json={"model": model, "prompt": PROMPT, "quality": quality},
        timeout=60,
    )
    r.raise_for_status()
    return time.perf_counter() - t if r.status_code == 200 else float("inf")

for model in ("openai/gpt-image-2.5", "openai/gpt-image-2.5-sunburst"):
    runs = [timed(model, "medium") for _ in range(10)]
    print(model, "median", round(statistics.median(runs), 1), "s", "slow:", runs.count(float("inf")))

Read the result carefully

Network distance, your own upload of references, and queueing on shared capacity all show up in a wall-clock timer. Compare models inside the same run and at the same hour. If Flare's median is clearly lower than Sunburst's, that matches the direction of the announcement; if it is not, your prompt may not be the everyday case OpenAI describes. Note that '50% faster' can be read as a time ratio of 0.5 or 0.67, so do not hold either number against your results.

Sync waits cap at 30 seconds on Sume's Image API. Treat any 202 as the right place to switch to the job pattern in Jobs and results.