Gemini 3.1 Deep Think

// Overview

Gemini 3.1 Deep Think is Google's specialized reasoning mode, dated February 2026 on its own model page. Google describes it as built on top of Gemini 3.1 Pro: rather than a separate base model, it is the configuration of that model that spends far more inference-time compute working a problem before it answers. Google positions it for science, research and engineering — the cases where being right matters more than being fast.

The published comparison puts it well ahead of the field on abstract and scientific reasoning. Against Gemini 3 Pro Preview Thinking (High), Opus 4.6 Thinking (Max) and GPT-5.2 Thinking (xhigh), it scores 84.6% on ARC-AGI-2 (ARC Prize verified) versus 31.1%, 68.8% and 52.9%; 48.4% on Humanity's Last Exam with no tools and 53.4% with search and code execution; and 81.5% on MMMU-Pro. On competition science it reports 81.5% on IMO 2025 mathematics, 87.7% on IPO 2025 physics theory, 82.8% on ICO 2025 chemistry theory, and 50.5% on the CMT-Benchmark for condensed-matter theory. Its Codeforces Elo without tools is 3455.

Google illustrates the mode with applied work rather than chat: optimising problems in materials science, catching errors in mathematical arguments, and prototyping in mechanical engineering. Access is through the Gemini app at gemini.google.com. Google's model page does not state a context window, knowledge cutoff or per-token pricing for this mode, so none is listed here.

// Benchmarks

This model's scores

1. ARC-AGI-2 (ARC Prize verified)84.6%
2. IPO 2025 (theory, physics)87.7%
3. ICO 2025 (theory, chemistry)82.8%
4. IMO 2025 (mathematics)81.5%
5. MMMU-Pro (no tools)81.5%
6. Humanity's Last Exam (search + code execution)53.4%
7. CMT-Benchmark (condensed matter theory)50.5%
8. Humanity's Last Exam (text + MM, no tools)48.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• 84.6% on ARC-AGI-2 (ARC Prize verified) — more than double the Gemini 3 Pro Preview Thinking (High) result of 31.1%
• 48.4% on Humanity's Last Exam with no tools, rising to 53.4% with search and code execution
• Competition-grade science: 81.5% on IMO 2025 mathematics, 87.7% on IPO 2025 physics theory, 82.8% on ICO 2025 chemistry theory
• 3455 Codeforces Elo with no tools, ahead of every peer in Google's published comparison
• 50.5% on the CMT-Benchmark for condensed-matter theory, the strongest score in the same table
• Built on Gemini 3.1 Pro, so it inherits that model's multimodal understanding (81.5% MMMU-Pro)

// Best for

• Research-level mathematics, including checking a proof or argument for errors
• Olympiad-grade physics and chemistry problems
• Materials-science and other scientific optimisation problems
• Mechanical-engineering prototyping and design reasoning
• Abstract reasoning tasks of the ARC-AGI-2 kind that defeat standard chat models
• Any task where extra inference-time compute is worth a slower answer

// Gemini Deep Think — every version

The full lineage of the Gemini Deep Think line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.