Shadow-run gpt-image-2.5 beside your current model before October 23

A shadow run sends each real prompt to the new model as well as the old one, and keeps only the new result for comparison. On Sume, give every shadow request its own Idempotency-Key, built from the order id and the model id. The same key with a different payload returns 409 idempotency_conflict, and a key shared across models would trigger exactly that.

Run it before OpenAI's gpt-image-1 shutdown on October 23, 2026, while the old model still answers.

What to record

• model as requested and usage.cost as billed USD from the response, so cost per model is measured, not guessed.
• The data[0].url for each model, so a person can open both.
• The HTTP status, since a slow job returns 202 and its cost arrives on the job result.

The shadow function

The function returns one row per model id. Wire the old model through your existing client, and only the Sume side is shown.

import json, os, urllib.request

def sume(model, prompt, key):
    req = urllib.request.Request(
        "https://api.sume.com/v1/images",
        json.dumps({"model": model, "prompt": prompt}).encode(),
        {"Authorization": "Bearer " + os.environ["SUME_API_KEY"],
         "Content-Type": "application/json", "Idempotency-Key": key})
    with urllib.request.urlopen(req, timeout=60) as r:
        return r.status, json.load(r)

def shadow(order_id, prompt, models):
    rows = []
    for model in models:
        status, body = sume(model, prompt, f"{order_id}:shadow:{model}")
        if status == 200:
            rows.append((model, body["data"][0]["url"], body["usage"]["cost"]))
        else:
            rows.append((model, "pending", None))
    return rows

Candidates worth shadowing

Pick a winner on your own prompts, then flip the config as in the model-map post. A traffic-percentage version of the same test is in the canary post.