GPT Image 2.5 Takes First Place in the November Model League

What did we test, and how did we evaluate it?

In the November 2026 Model League, we compared six image-generation models across five architectural prompts. The prompts covered a residential exterior at twilight, a Scandinavian living room, a stone villa and pool in Bodrum, a kitchen material detail, and an aerial view of an urban square. Each model was tested through the same API using its standard settings.

The images were evaluated by an AI judge that did not know the model names. The scoring weighted prompt adherence at 30%, architectural accuracy at 30%, photorealism at 25%, and overall polish at 15%. Scores were given out of 100. The times and costs below show the average generation time per model and the cost per image.

Rankings

GPT Image 2.5 earned the highest overall score and also had the lowest cost per image in the table. Seedream 5.0 Pro came in second, but had the longest average generation time on the list. FLUX.2 [pro] matched GPT Image 2.5 for speed, with an average generation time of 10.9 seconds.

Standouts by prompt

Residential exterior: GPT Image 2.5 won with a score of 90. The judge found the two-storey concrete-and-wood structure, large windows, flat eaves, and gravel landscaping with olive trees clearly recognizable. The balance between blue hour and warm interior lighting was also rated positively. FLUX.2 [pro] rendered the facade lines and lighting naturally, though its window openings were more limited.

Scandinavian living room: Seedream 5.0 Pro took first place with a score of 90. The window light, two plants, perspective, and materials were judged natural, with no notable flaws reported. In the other examples, extra decor, furniture, or framing choices detracted from the simple layout described in the prompt.

Stone villa and pool in Bodrum: GPT Image 2.5 scored 95.5, the highest prompt result in the test. It captured the stone villa with large windows, bougainvillea, infinity pool, and sunset light effectively, with no significant distortions in perspective or materials.

Kitchen material detail: Seedream 5.0 Pro stood out with a score of 93. The faucet, countertop, and cabinet geometry were consistent, while the soft lighting and product-photography feel were successful. GPT Image 2.5 used a wider framing, which reduced the impact of the requested close-up detail.

Aerial view of an urban square: Nano Banana 2 won with a score of 90. The dense urban fabric, brick-and-glass buildings, water feature, and tram were all strongly represented. GPT Image 2.5’s result also matched the prompt closely, with a grid of trees, a tram, and pedestrians shown at scale.

Which model is right for which job?

If you’re weighing overall score, speed, and cost together, GPT Image 2.5 stands out: it had the highest average score and the lowest cost per image in this test. Its results for the residential exterior and Bodrum villa prompts in particular make it a strong option for generating architectural scenes.

Seedream 5.0 Pro’s two wins are worth noting for interior and materials-focused work, but factor in its longer generation time and higher cost. Nano Banana 2 stood out for its performance on the complex urban square prompt, though it finished sixth overall. If speed is the priority, FLUX.2 [pro] matched GPT Image 2.5’s average time, but scored lower overall.

Limitations

We generated each prompt only once, so results may vary across repeated attempts. The evaluation was conducted by a blind AI judge, not by humans. Results also depend on the language and style of the prompts used: a different prompt format could change the rankings or performance on individual prompts. The scores are therefore best understood as a comparison under these test conditions, rather than as an absolute measure of quality.

See all the images side by side on the Model League page.