Resolves as YES if OpenAI releases a model which generates images using a surrogate reasoning process in a manner that scales with respect to inference-time compute, before January 1st 2026.

In order for this question to resolve as YES, the method must be able to perform the image generation from a single prompt interaction with the user, and use a variable amount of compute as a function of the difficulty of the image generation task (similar to o1).

Methods which perform the reasoning steps using text tokens are acceptable, provided that the technique produces intermediate images during the image generation process (without requiring user input) -- as opposed to simply reasoning about the the optimal prompt and then generating the image from that. Full intermediate images are not a requirement, the model can generate intermediate patches for this to qualify.

Methods which perform the reasoning steps using additional image tokens, a visual scratchpad, or which superimpose the reasoning process on the image itself (by e.g. adjusting the image token vectors to highlight regions) are also acceptable.