AI Models Navi
모델 목록으로
Google Deep Mind독점

Gemini 2.5 Deep Think

Gemini 2.5 Deep Think is an inference model developed by Google DeepMind. It is a chat-specialized foundational model equipped with a very long context window of 1000K.

파라미터

Undisclosed

컨텍스트

1000K

라이선스

Proprietary

출시일

2025-08-01

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

  • ・Extremely long context window.
  • ・Achieves advanced reasoning capabilities.
  • ・Developed by Google DeepMind.

약점

  • ・Closed-source license.
  • ・Internal model details are undisclosed.
  • ・Potential usage restrictions.

활용 사례

  • ・Analysis of large documents.
  • ・Complex logical reasoning tasks.
  • ・Advanced conversational chat use.

심층 분석

GPQA Diamond

92%

PhD-level science reasoning, tops Claude 4.7 (88%) and GPT-5 (90%)

SWE-bench Verified

~78%

Coding benchmark, behind Claude 4.7 (~85%) and GPT-5 (~80%)

Arena Elo

~1465

Based on Gemini 2.5 Pro base, Deep Think adds extended reasoning

Context Window

1M tokens

1,000,000 input tokens

Input Price

$1.25/M tokens

Cheapest frontier reasoning model at 1M-token scale

Output Price

$10.00/M tokens

Standard output pricing for reasoning tier

Output Speed

~30 tok/s

Slower than GPT-5 (~110) and Claude 4.7 (~80) due to deep reasoning chains

Release

April 2026 (GA)

Opened to all paid API users, previously AI Ultra only

강점

  • ・Highest GPQA Diamond score (92%) among frontier models as of May 2026
  • ・Cheapest frontier reasoning model at $1.25/M input
  • ・1M token context with 98% needle-in-haystack at 500K
  • ・Gold-medal level on International Math Olympiad
  • ・Strong needle-in-haystack retrieval (98% at 500K tokens)

약점

  • ・Slower output speed (~30 tok/s) vs competitors due to extended thinking
  • ・Time-to-first-token ~1.1s, slower than GPT-5 (0.4s) and Claude 4.7 (0.6s)
  • ・Weaker coding performance (78% SWE-bench) vs Claude 4.7 (85%)
  • ・Computer use / browser capability still experimental
  • ・Deep Think mode requires more compute, increasing effective cost

경쟁사 비교

ModelArenaSWEGPQAPrice
Claude 4.7 Sonnet~1470~85%88%$3/$15 per 1M
GPT-5~1480~80%90%$5/$20 per 1M
Gemini 2.5 Deep Think~1465~78%92%$1.25/$10 per 1M

Gemini 2.5 Deep Think is Google's extended-reasoning mode built on Gemini 2.5 Pro, opened to all paid API users in April 2026. It leads on pure reasoning benchmarks (GPQA 92%, ARC-AGI) but trails Claude 4.7 and GPT-5 on coding tasks. Its standout advantage is the lowest pricing among frontier reasoning models at $1.25/M input tokens with a 1M context window.

출처

분석 생성일: 2026-05-24