When generation learned to reason

Through 2025 the recipe was fixed: encode the promptPromptThe text description you provide to a model to specify what you want it to generate., denoise a latent, decode to pixels. In 2026 a different idea started to matter, that a model might think before it generates, the way a person sketches and plans before committing paint.

GPT-Image-2, released by OpenAI in early 2026 as ChatGPT Images 2.0, is the clearest example. In its Thinking mode, available to paid tiers, the model plans a layout, can search the web for references, and checks its own output against the prompt before returning it. OpenAI has not disclosed whether it is diffusion, autoregressiveAutoregressiveGenerating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion., or a hybrid, describing it only as a generalist model, so treat any confident architecture claim with suspicion. What is verifiable is the behavior and the result: it spends inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training.-time compute reasoning about the image, and it took the top of the Artificial Analysis image arena by the largest margin that board has recorded.

This is the move that reshaped language models, inference-time reasoning, arriving in images. It trades latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput. and money for adherence and correctness, which is exactly the trade a professional wants on a hard brief and a bad one for a throwaway image.

Underneath the products, a quieter architectural shift runs in parallel. A wave of 2026 research revisits autoregressive image generation, models that predict an image as a sequence of tokens the way an LLM predicts words, rather than denoisingDenoisingThe core loop of diffusion. Starting from pure noise, the model repeatedly predicts and subtracts a little noise until a coherent image emerges. all at once. Work such as OmniGen-AR and a line of randomized-parallel and frequency-autoregressive methods suggests diffusion's hold on high-end image generation is no longer total. It is too early to call a winner; the point is that the generative mechanism itself is back in play.

For an operator the practical signal is simple. The best results on hard, specification-heavy prompts now come from models that are slower and more expensive per image because they are thinking, and that cost is a feature, not a regression.