Google releases Gemini 2.5 Deep Think for AI Ultra subscribers
Briefly

Google has launched its Gemini 2.5 Deep Think model, focused on complex queries and requiring substantial compute resources. Accessibility is limited to subscribers of the $250 AI Ultra plan. Deep Think enhances the foundational capabilities of Gemini 2.5 Pro by increasing "thinking time" and enabling more parallel analysis. This model revisits and remixes hypotheses for improved output quality. It excels in tasks requiring design aesthetics, scientific reasoning, and coding, taking several minutes for responses. It significantly outperforms standard benchmarks, scoring 34.8 percent on Humanity's Last Exam, compared to competitors that peak at 20-25 percent.
Deep Think is designed for the most complex queries, which means it uses more compute resources than other models.
Deep Think explores multiple approaches to a problem, even revisiting and remixing the various hypotheses it generates.
Deep Think takes several minutes to come up with an answer, making it more adept at design aesthetics, scientific reasoning, and coding.
Gemini 2.5 Deep Think managed a score of 34.8 percent on Humanity's Last Exam, surpassing other models which top out at 20 or 25 percent.
Read at Ars Technica
[
|
]