Gemini Deep Think
#35 in Reasoning Modelsgoogle · since Erstmals vorgestellt/verfügbar für Google AI Ultra-Abonnenten am 4. Dezember 2025 (Gemini 3 Deep Think); größeres Upgrad · 10× · last seen Jun 30, 2026
Gemini Deep Think is a specialized reasoning mode from Google DeepMind built on top of the Gemini 3.1 Pro model, designed for complex, multi-step problems in mathematics, science and programming. Instead of a single linear reasoning path, it uses parallel reasoning that explores and evaluates multiple hypotheses simultaneously. The model achieved gold-medal-level performance at the 2025 International Mathematical Olympiad as well as the written sections of the 2025 Physics and Chemistry Olympiads, and tops benchmarks such as Humanity's Last Exam and ARC-AGI-2. Access is limited to Google AI Ultra subscribers within the Gemini app, with an additional API early-access program for select researchers, engineers and enterprises.
Features
| Key Benchmark (%) | Humanity's Last Exam: 48.4% (without tools); ARC-AGI-2: 84.6%; Codeforces Elo: 3455; gold-medal level at IMO 2025 and Physics/Chemistry Olympiad 2025 (written parts) |
| Context Window (Tokens) | Up to 1 million tokens (Deep Think variant based on Gemini 3.1 Pro architecture) |
| License | Proprietary; usage under Google Cloud Platform terms of service or Google AI Ultra subscription (not open source) |
| Multimodality | Text, image, audio, video, and code as input (based on Gemini 3.1 Pro); output primarily text |
| Platform | Gemini App (Google AI Ultra); Gemini API/Vertex AI (early access program for researchers, developers, enterprises) |
| Price per 1M Tokens | No official separate API price published; consumer access only via Google AI Ultra starting at $99.99/month; comparable Gemini 3.1 Pro costs $2 (input) / $12 (output) per 1M tokens |
| Release Date | Gemini 3 Deep Think: December 4, 2025 (rollout to Ultra subscribers); upgrade Gemini 3.1 Deep Think: February 2026 |