Google DeepMind has upgraded its specialized reasoning mode, Gemini 3 Deep Think, and is making it available through the Gemini app as well as via an API in an early-access program on Vertex AI. The upgrade is designed to support complex tasks in science, research, and engineering. The Gemini app is available to Google AI Ultra subscribers, while developers and researchers can apply for access to the API program.
According to Google DeepMind, Deep Think achieves state-of-the-art results across several benchmarks: ARC-AGI-2 (a test of logical reasoning), Humanity’s Last Exam (challenging problems in mathematics, science, and engineering), and an Elo rating of 3,455 on the competitive programming platform Codeforces.
|
Benchmark |
Deep Think |
Claude Opus 4.6 |
GPT-5.2 |
Gemini 3 Pro Preview |
|---|---|---|---|---|
|
ARC-AGI-2 |
84.6% |
68.8% |
52.9% |
31.1% |
|
Humanity’s Last Exam |
48.4% |
40.0% |
34.5% |
37.5% |
|
Codeforces (Elo) |
3,455 |
2,352 |
– |
2,512 |
In addition, the model achieved gold-medal–level performance at the 2025 Physics and Chemistry Olympiads. Google DeepMind has also shared examples demonstrating the use of Deep Think in scientific research applications.
ES 
EN