Google DeepMind has introduced Gemini 3 Deep Think, a specialized version of Gemini 3 Pro designed to tackle complex challenges in scientific research, engineering, and advanced mathematics. The new mode is available through the Google AI Ultra subscription, priced at approximately $250 per month.
The goal is to turn the model into a true “scientific companion,” capable of spending more time reasoning through difficult problems rather than prioritizing speed. Deep Think is optimized to allocate more compute to multi-step reasoning, producing deeper analytical outputs than standard models.
According to Google, Gemini 3 Pro delivers more than a 50% improvement over Gemini 2.5 Pro in advanced reasoning evaluations. On Humanity’s Last Exam, a benchmark designed to test frontier models with extremely advanced academic-level questions, Gemini 3 Deep Think reportedly achieves 48.4% without external tools, among the highest scores reported so far.
Deep Think is designed for high-complexity use cases, including:
advanced mathematical proofs
scientific reasoning across disciplines
identifying logical errors in research
engineering problem solving
long-form multi-step analysis
Unlike standard models, Deep Think is intentionally slower, trading latency for depth. This reflects a broader shift in frontier AI: systems designed to “think longer” in order to solve harder problems.
Google also stated it is working directly with scientists and researchers to apply Deep Think to previously intractable problems, opening new possibilities for AI-accelerated discovery.
At the same time, the company has already previewed Gemini 3.1 Pro, signaling rapid iteration in this new class of reasoning-focused AI systems.
Gemini 3 Deep Think marks a shift: AI models are no longer just getting faster, they’re learning to think deeper.
