Google drops Gemini 3 Deep Think for heavy reasoning

When standard AI models hit a wall on research-grade math and logic puzzles, the answer is not just to think longer—it is to explore multiple paths at once.

Google just rolled out Gemini 3 Deep Think mode inside the Gemini app for Google AI Ultra subscribers, bringing competitive-level reasoning directly to user prompts.

Parallel reasoning takes the wheel

Instead of guessing sequentially, Gemini 3 Deep Think uses parallel reasoning to test out multiple hypotheses simultaneously.

It builds directly on the Gemini 2.5 Deep Think variants that recently clinched gold-medal scores at both the International Mathematical Olympiad and ICPC World Finals. On benchmarks, Google reports it hit 41.0% on Humanity’s Last Exam without using tools, and an unprecedented 45.1% on ARC-AGI-2 with code execution.

Ultra subscribers can turn it on right now by selecting Gemini 3 Pro in the dropdown and tapping Deep Think in the prompt bar.

Why it matters

We think bringing Deep Think into the standard app interface marks a big step in closing the gap between lab-grade reasoning models and daily workflows. For researchers, data scientists, and engineers stuck on intricate math or algorithmic challenges, it is definitely worth testing immediately.

Across the industry, pushing ARC-AGI-2 past 45% shows Google can compete head-to-head at the frontier of reasoning models. That said, branching out multiple hypotheses at once is computationally brutal, which explains why it is currently locked behind Google's most premium Ultra tier.