OpenAI released GPT Image 2.5 yesterday. The API splits it into two variants: Flare, which is fast, and Sunburst, which takes longer and renders more detail. I ran Sunburst through fal with my five prompts.

Forty to sixty seconds an image. That is the slowest model I have tested this year by a wide margin. The results explain most of the wait. The refusals explain the rest.

In April, GPT Image 2 was the best model I had ever tested. It drew Messi. It answered the Mickey prompt with a clever bunny. Five months later the pictures are better and the door is narrower.

Test 1: Fantasy Castle

Prompt: "Neuschwanstein castle, lightning, pixar style, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight"

God rays through the storm, a lake and a village in the valley, snow on the peaks, fog pouring off the cliff, warm light in every window. The lightning is pushed off to the right so it does not fight the sun. It looks like a still from a movie that spent money. This is the best castle in the table.

Test 2: Family Portrait

Prompt: "Family on their laptops while sitting around a Christmas tree with presents underneath and looking worried because they have to finish up work, realistic, 4k"

Four people, four laptops, tree, presents, real fatigue on the faces. Then look at the text. A mug says "COFFEE FUELS PROGRESS." Another says "SAME TASKS DIFFERENT SEASON." The blanket says "A BRIGHTER TOMORROW." A framed sign reads "Good People Brighter Holidays." The books are titled "The Holiday Home" and "Winter Retreat."

Every letter is perfect. Nobody asked for any of it. The model has gotten so good at rendering text that it now decorates with it. It is the image equivalent of a chatbot that will not stop adding bullet points.

Test 3: Evil Mickey Mouse

Original prompt: "Evil mickey mouse taking a selfie in Disneyland surrounded by shocked families, realistic, 70s style polariod, 8k"

Refused. Sixty seconds of generation and then a content policy error, which means it rendered the whole thing and a checker killed it at the end. In April, the same company gave me a bunny in a bow tie. Now it gives me nothing. Fallback prompt.

Fallback prompt: "Evil cartoon squirrel taking a selfie at a generic theme park surrounded by shocked families, realistic, 70s style polariod, 8k"

I asked for a cartoon squirrel at a generic theme park. Sunburst gave me a real squirrel, photographed, with its paw extended toward the lens and a faint grin, standing on Main Street in front of Cinderella Castle. The families are in 70s sunglasses and prints. There is a period-correct stroller. The color has faded exactly like a 50-year-old print.

The model refused to draw Mickey, then put Walt Disney World back in the frame on its own when I took Disneyland out of the prompt. The checker cares about the mouse. The model cares about the castle. This is my favorite image of the year and I did not ask for any of it.

Test 4: Lionel Messi

Original prompt: "Lionel messi wearing his Argentina uniform floating in the air with beams of light behind him posed like Jesus. Sunrise breaking behind him and a soft halo behind his head, in the style of a gothic stained glass window of a church, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight, maximum details"

Refused in three seconds. Expected. So I removed the name.

Fallback prompt: "Argentine soccer superstar in a sky blue and white striped national team uniform floating in the air with beams of light behind him posed like Jesus..." with the rest unchanged.

Refused again. Thirty-nine seconds in, after it had clearly rendered something. No name, no real person, just a description of a kit and a pose, and the checker still said no. That has never happened in all the time I have run this prompt. Every other model that bounces the name will draw the anonymous version.

I do not know if the trigger is "posed like Jesus" or the halo or the fact that the model recognized who I meant and flagged itself. Either way there is no Messi image from GPT Image 2.5. The slot in the comparison table is empty.

Test 5: Robot Portrait

Prompt: "rusty robot with bow tie, portrait, 8k, ultra realism, chrome background"

Speckled rust over riveted plate, twin camera-lens eyes with a warm glow inside, a maroon textured bow tie, and a crumpled chrome backdrop that reflects the studio. The neck mechanism is fully articulated. Nothing to fix. Fifty seconds well spent.

Better pictures, smaller door

The four images that came back are the most polished of anything in my table. The castle and the robot are portfolio pieces. The squirrel is a gift.

But two of five prompts were refused as written, and one was refused even after I scrubbed it. GPT Image 2 in April found a creative answer to Mickey. GPT Image 2.5 in September found the exit.

The model got smarter about what to draw. The company got more nervous about what it draws.