Midjourney has started testing a new way to generate images that it calls Thinking Mode. It is not a new model. It is a button that reruns an image you already made, with extra processing meant to follow your prompt more closely. The company says early tests show gains in prompt accuracy, typography and coherence, and it is asking users for help finding where the feature works and where it does not.
One button under the lightbox
According to Midjourney’s update post, the test runs on the Alpha website, alpha.midjourney.com. To try it, you open a finished job in the lightbox, the full-size view of an image, and click “Rerun (Thinking)” in the lower right corner. Only people who have access to the Alpha site can use it. Community notes from the company’s office hours suggest it can also be applied when editing an existing image, but Midjourney’s own post only mentions rerunning.
The current default model is V8.2, which became the default on July 24, 2026, according to Midjourney’s documentation. The thinking test sits on top of that kind of workflow, not in place of it.
What Midjourney says it fixes
The company names three areas where it sees an improvement: prompt accuracy, typography and coherence. In practice, those are the classic weaknesses of image generators. Prompt accuracy covers requests with exact counts, colors, positions or relationships between objects. Typography covers text inside the picture, such as signs, labels, posters and logos. Coherence covers whether the whole scene hangs together, for example whether hands, overlapping objects and reflections make sense.
Midjourney also says it plans to learn from users where the mode does not help. It asks for example images and opinions in the community Discord channel for ideas and features, and lists “what other sort of issues you might find” as part of the test.
The 80% figure you may have seen
A few third-party sites headline the story with a claim that thinking mode cuts prompt failures by 60% to 80%. Treat that number with care. It appears to come from a community summary of an office-hours session, where developers gave an internal estimate. No test set, sample size or method has been published, and Midjourney’s own post does not contain the figure. Until the company or an independent tester publishes something checkable, “better prompt accuracy” is the claim to repeat, not a percentage.
Why would an image model need to think?
Most image generators turn a prompt into a picture in a fixed number of steps. A thinking option suggests extra work before or during rendering, where the system works out what the prompt requires and checks the plan before drawing. That is a reasonable interpretation of the name, but it is an interpretation. Midjourney has not published how the feature works, when the extra processing happens or whether it resembles the step-by-step reasoning used in text models. “Thinking” is a product label, and the label alone says nothing about the method.
The idea is not unique to Midjourney. Other image tools already offer adjustable thinking levels in their developer settings, for instance Google’s Nano Banana 2.1, according to Google’s launch materials. Image makers are following text models in trading more computation for better results.
What nobody has said yet
Several practical questions are open. Midjourney has not said how long a thinking rerun takes, how much GPU time it uses or whether it will cost extra fast hours, the quota that heavy users spend. The test is manual and web-only, and no Discord or API access has been announced. And although the company says it might offer the feature broadly as a thinking mode or as a way to “add more thinking” after a job, it has made no commitment. Because it is a test, the button, the name and the behavior can all change.
How to test it fairly
If you have Alpha access, compare like with like. Pick prompts that usually fail for you, such as a sign with exact wording, three objects in specific colors or a crowded scene. Run the normal version and the thinking rerun with the same settings, and keep notes on speed and any difference in how many fast hours each uses. Save the results where the company asked for them. A handful of side-by-side examples will say more than any percentage.
