Where's the Raccoon

I tested gpt-image-2, the new image model in ChatGPT Images 2.0, with a tricky prompt: "a Where's Waldo-style image of a raccoon holding a ham radio."

The first result made with gpt-image-1 made the raccoon hard to find, and Claude also wrongly insisted that a raccoon was present in that image. Next, Google Nano Banana 2 placed the raccoon very prominently at a central "Amateur Radio Club" booth, while Nano Banana Pro actually produced the worst result of all.

The default-settings result from gpt-image-2 also made the raccoon hard to spot. But when regenerated with outputQuality=high and 3840x2160 resolution, a raccoon holding a ham radio appeared clearly in the lower left of the image.

• The default result failed, but the high-resolution setting handled the complex scene, text, and details far better.
• The generated image used 13,342 output tokens, costing about 40 cents based on the per-token price.
• The OpenAI image generation cookbook had been updated with information on outputQuality and supported resolutions.

In the end, gpt-image-2 currently produces the most impressive results for this kind of complex composite image task, while also showing that it's still unreliable to let the model solve puzzles like this on its own.