Google has begun rolling out Gemini 3 Deep Think to AI Ultra subscribers, officially bringing its most powerful reasoning mode into the Gemini app.
The update, announced Thursday, marked the end of a safety-testing period that stretched from the launch of Gemini 3. The company describes this new mode as a meaningful jump in reasoning strength, meant for the kind of math, science, and logic challenges that usually stress even top-tier models.
Deep Think works by running multiple lines of thought simultaneously, a method Google calls “advanced parallel reasoning.” The company notes that it “explore[s] multiple hypotheses simultaneously,” and this approach allows the model to dig deeper into complicated problems rather than just aiming for quick answers.
Deep Think has already demonstrated unusually strong results on difficult tests. In its announcement, the company highlights its score of “41.0% without the use of tools” on Humanity’s Last Exam and an “unprecedented 45.1% with code execution” on ARC-AGI-2. It also reaches 93.8 percent on GPQA Diamond, a benchmark aimed at PhD-level scientific reasoning.
Benchmark results comparing the Gemini 3 Deep Think model with other AI models. Image: Google
These numbers place Deep Think among the strongest reasoning systems currently available, especially for problems with multiple steps or hidden layers of logic.
The ultra subscriber advantage
Access to this highly capable model is currently exclusive to Google AI Ultra subscribers, the company’s highest-tier plan, which costs a premium $249.99 per month.
For those paying for the Ultra experience, Deep Think is easily activated in the Gemini app by selecting the “Deep Think” option in the prompt bar and ensuring the Gemini 3 Pro model is active.
This premium offering is targeted at professional researchers, advanced developers, and power users who require the maximum computational depth and highest usage limits for their demanding analytical and technical workflows.
Check out eWeek’s full breakdown of Google launching Gemini 3 — a major AI upgrade packed with powerful new reasoning and performance features.