Midjourney began testing Thinking Mode in an official update on October 8, 2026. For now it exists only on the Alpha website: open a finished job and click Rerun (Thinking) to regenerate it with an extra round of thought before the image is produced, with the goal of fixing things the prompt specified but the picture got wrong. An image model borrowing the "thinking" playbook from language models is the signal worth noting in this small update. Prompt adherence is exactly where leading image tools separate themselves, as an earlier comparison of how to choose among major AI image tools showed.
It changes the rerun, not the model
The entry point is light: under a job's lightbox, in the lower right corner, sits Rerun (Thinking). In its testing note, Midjourney says internal tests show the mode helping with prompt accuracy, typography in images, and coherence, and it invites users to post cases where it does and does not help in a designated Discord channel. The company also sketches the direction: in the future this might become a broadly available thinking mode, or a way to add more thinking after a job completes. In other words, this is an experiment attached to a rerun button, not a new model release, and no timeline for a general launch has been given.
Why image models are starting to "think" too
The language-model lesson is already visible: spending more compute buys stronger instruction following. Image generation's classic failures sit precisely there — wrong element counts, misspelled text, tangled relationships between subjects — essentially drawing before thinking. A rerun with a thinking pass turns a one-shot draw into a redo with review built in. For creators, the practical effect is procedural: instead of discarding a whole batch, you can apply one corrective pass to the image closest to what you wanted, which changes how the cost of trial and error is distributed.
Who can try it now, and how high to set expectations
The audience is narrow: Alpha website users only, who must compare before-and-after results themselves. Midjourney has published no quantitative metrics and no word on the rerun's time or cost, so claims about how much it improves cannot be verified yet and should not be treated as proven capability. Images with many elements, embedded text or complex subject relationships are the best test cases, since they expose adherence problems most clearly; for simple compositions, a thinking rerun will likely add little. The safest reading is that this is a direction under validation — whether it holds up will be decided once it leaves Alpha with reproducible comparisons.
