Gemini 3 Deep Think February 2026 Upgrade
Gemini 3 Deep Think 2026년 2월 업그레이드는 Google DeepMind에서 개발한 추론 모델입니다. 1M(100만) 토큰의 대규모 컨텍스트 윈도우를 특징으로 하며, 고급 추론 기능을 제공합니다.
파라미터
Undisclosed
컨텍스트
1M
라이선스
Proprietary
출시일
2026-02-13
API 가격
이 모델의 API 가격 정보는 현재 공개되지 않았습니다
강점
- ・강력한 추론 능력 탑재
- ・100만 토큰의 긴 컨텍스트
- ・Google DeepMind 개발
약점
- ・오픈 소스가 아닌 라이선스
- ・상세 벤치마크 미공개
- ・폐쇄적 사용 시스템
활용 사례
- ・복잡한 논리적 사고가 필요한 작업
- ・초장문 문서 분석
- ・고급 문제 해결에 적용
심층 분석
ARC-AGI-2
84.6%
ARC Prize verified, 15.8pp above Claude Opus 4.6 (68.8%), 31.7pp above GPT-5.2 (52.9%)
GPQA Diamond
93.8%
PhD-level science, slightly above GPT-5.2 (93.2%) and Claude Opus 4.6 (91.3%)
Codeforces Elo
3455
Legendary Grandmaster status, far above Claude Opus 4.6 (2352)
Humanity's Last Exam
48.4% (no tools)
New standard; 53.4% with search + code execution
IMO 2025
81.5%
Gold-medal level performance on International Math Olympiad
Context Window
1M tokens
1,000,000 input / 64,000 output
Input Price
$2.00/M tokens
$4.00/M for prompts >200K tokens
Output Price
$12.00/M tokens
$18.00/M for prompts >200K tokens
Release Date
February 12, 2026
Major upgrade to Gemini 3 Deep Think reasoning mode
강점
- ・Undisputed leader on abstract reasoning (ARC-AGI-2 84.6%) and competitive programming (Codeforces 3455)
- ・Gold-medal performance on IMO, IPhO (87.7%), and IChO (82.8%) 2025
- ・Strongest scientific reasoning across chemistry, physics, and condensed matter theory
- ・Multimodal input support (text, images, audio, video)
- ・1M token context window
약점
- ・Trails Claude Opus 4.6 on agentic enterprise tasks (GDPval-AA ~1200 vs 1606)
- ・Weaker on practical coding (SWE-bench 76.2% vs Claude 80.8%)
- ・Higher latency due to deep reasoning chains
- ・Higher cost than Gemini 2.5 Deep Think ($2/$12 vs $1.25/$10)
- ・Early API access only (not broadly available as of Feb 2026)
경쟁사 비교
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| Gemini 3 Deep Think | ~1500 (est) | 76.2% | 93.8% | $2/$12 per 1M |
| Claude Opus 4.6 Thinking Max | ~1490 | 80.8% | 91.3% | $15/$75 per 1M |
| GPT-5.2 Thinking xhigh | ~1480 | 80.0% | 93.2% | $5/$20 per 1M |
| Gemini 3 Pro (standard) | ~1470 | 76.2% | 91.9% | $2/$12 per 1M |
The February 2026 upgrade to Gemini 3 Deep Think is Google's most powerful reasoning mode, achieving state-of-the-art results on ARC-AGI-2 (84.6%), Codeforces (3455 Elo), and multiple international science olympiads. It excels at abstract reasoning, mathematical proofs, and scientific analysis but trails behind Claude Opus 4.6 on practical agentic and enterprise tasks.
출처
분석 생성일: 2026-05-24