Robots Atlas>ROBOTS ATLAS
Gemini 3.1 Deep Think
AI Models›Gemini

Gemini 3.1 Deep Think

3.1 Deep Think · Family: Gemini
Gemini 3.1 Deep Think is a specialized reasoning mode by Google DeepMind built on top of Gemini 3.1 Pro, designed for the most challenging scientific, mathematical, and engineering problems.
⏳ Preview⏳ Limited accessLLMMultimodalReasoning model📁 Gemini
Context window
1M
tokens
Max output
65,536
tokens
Release date
1 February 2026
🔬Google DeepMindResearch lab🏢GoogleOwner
Access:HostedDeployment:☁ Cloud

Overview

Gemini 3.1 Deep Think is a specialized reasoning mode developed by Google DeepMind, built on top of Gemini 3.1 Pro. The previous version (Gemini 3 Deep Think) was available from February 2026. Deep Think mode is designed for the most challenging scientific, mathematical, and engineering problems.

The model achieved 84.6% on ARC-AGI-2 (ARC Prize verified), 81.5% on IMO 2025, and 87.7% on IPhO 2025. Available via the Gemini app (Deep Think mode) for Gemini Ultra subscribers.

DeepMind's page presents results for 'Gemini 3.1 Deep Think Feb 2026'. Model parameters have not been disclosed. No separate model card is publicly available.

Classification
LLMMultimodalReasoning model
Family: Gemini
Applications
Research assistanceCodingData analysisQ&A / Question answering
Access & deployment
Hosted
Cloud
Weights: Closed
Key parameters
📏 Context: 1M
📥 Input: text, image

Technical specification

Context window
1M
tokens
Max output tokens
65,536
tokens per response
Knowledge cutoff
1 Jan 2025
Knowledge boundary
License
proprietary
Hardware requirements
Available via the Gemini app (Gemini Ultra subscription).
Modalities
⬇ Input
textimage
⬆ Output
text

Capabilities and applications

Native model capabilities
Reasoning
Category: reasoning
Multi-step reasoning
Category: reasoning
Multimodal understanding
Category: multimodal
Long context
Category: reasoning
Image understanding
Category: vision
Diagram reasoning
Category: reasoning
Application domains
Research assistanceCodingData analysisQ&A / Question answering

Benchmark results

9 benchmarks
ARC-AGI-2
accuracy · ARC Prize Verified, Gemini 3.1 Deep Think Feb 2026
84.6%%
📄 https://deepmind.google/models/gemini/deep-think/
Abstract reasoning puzzles, ARC Prize verified.
Humanity's Last Exam
accuracy · No tools, Gemini 3.1 Deep Think Feb 2026
48.4%%
📄 https://deepmind.google/models/gemini/deep-think/
Full set (text + MM). No tools.
Humanity's Last Exam
accuracy · Search + code execution, Gemini 3.1 Deep Think Feb 2026
53.4%%
📄 https://deepmind.google/models/gemini/deep-think/
Full set (text + MM). With search and code execution.
MMMU-Pro
accuracy · No tools, Gemini 3.1 Deep Think Feb 2026
81.5%%
📄 https://deepmind.google/models/gemini/deep-think/
Multimodal understanding and reasoning.
International Math Olympiad 2025
accuracy · Gemini 3.1 Deep Think Feb 2026
81.5%%
📄 https://deepmind.google/models/gemini/deep-think/
Mathematics – International Mathematical Olympiad 2025.
Codeforces
elo · No tools, Gemini 3.1 Deep Think Feb 2026
3455 EloElo
📄 https://deepmind.google/models/gemini/deep-think/
Coding and algorithms, no tools.
International Physics Olympiad 2025 (theory)
accuracy · Gemini 3.1 Deep Think Feb 2026
87.7%%
📄 https://deepmind.google/models/gemini/deep-think/
Physics – International Physics Olympiad 2025 (theory).
CMT-Benchmark
accuracy · Gemini 3.1 Deep Think Feb 2026
50.5%%
📄 https://deepmind.google/models/gemini/deep-think/
Condensed matter theory.
International Chemistry Olympiad 2025 (theory)
accuracy · Gemini 3.1 Deep Think Feb 2026
82.8%%
📄 https://deepmind.google/models/gemini/deep-think/
Chemistry – International Chemistry Olympiad 2025 (theory).

Technical architecture

Sources and related pages