$ cat wiki/models/gemini-3-1-deep-think.md

Gemini 3.1 Deep Think

modelupdated 2026-07-21created 2026-05-17

Spec

Attribute	Value
Developer	Google DeepMind
Released	unknown
Announced	2026-01 (advanced version; original Deep Think achieved IMO Gold July 2025)
Context window	unknown
Pricing	unknown
License	unknown
Availability	unknown
Model family	Gemini 3.x
Variant	Deep Think (extended reasoning mode)
Focus	Mathematical & scientific reasoning, autonomous research
IMO-ProofBench Advanced	up to 90% (advanced version, Jan 2026)
Prior milestone	IMO Gold-medal standard (July 2025)

IMO-ProofBench Advanced: up to 90%
2025 International Mathematical Olympiad: Gold-medal level (natural language reasoning)
2025 International Physics Olympiad (written): Gold-medal level
2025 International Chemistry Olympiad (written): Gold-medal level

Google DeepMind built an autonomous research agent codenamed Aletheia powered by Gemini Deep Think:

Natural language verifier to identify proof flaws
Iterative generate-revise loop; can admit unsolvability
18 previously unsolved problems cracked across math, physics, CS, economics
Disproved a decade-old conjecture (online submodular optimization, unsolved since 2015)
Generated paper Feng26 with no human intervention — calculating eigenweights in arithmetic geometry → One of the first instances of genuine autonomous mathematical research by AI

Beyond math: excels across chemistry and physics at research level.

Google DeepMind
AlphaEvolve — companion coding/algorithm agent also powered by Gemini
Reasoning Models — Gemini Deep Think is a leading instance of test-time compute scaling
Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling — parallel result from a different group, same capability frontier