Introducing Gemini 2.5 Deep ThinkGoogleReleased August 1, 2025

Gemini 2.5 Deep Think

Google’s dedicated reasoning model utilizing test-time compute to solve Olympiad-level mathematics and competitive coding.

reasoning-llmProprietary API$2.00 / 1M tok (with Thinking)Context: 1.049M (1,048,576 tokens)Arena ELO: 1385

Technical Specifications

Architecture Type
Test-Time Compute Reasoning Transformer
Total Parameters
Undisclosed
Context Window
1.049M (1,048,576 tokens)
Max Output Tokens
65.536K (65,536 tokens)
Knowledge Cutoff
March 2025
Supported Modalities
text, code, reasoning
License & Access
Google Cloud API Terms of Service

Benchmark Evaluations

Math500
94.8
Aime 2024
86.7
Gpqa Diamond
72.1
Codeforces ELO
2280
Chatbot Arena ELO
1385

Deep Architectural Overview

Gemini 2.5 Deep Think dynamically scales inference-time reasoning steps before emitting final answers. It excels in verified theorem proving, scientific research, and complex system debugging.

Strengths & Considerations

Core Strengths
  • World-class AIME & Olympiad math scores (86.7%)
  • Transparent step-by-step thinking trace
  • High resistance to hallucinations in logical proofs
Known Limitations
  • Longer latency due to extended thinking loops

Token & API Pricing

Input Tokens (1M)$2.00
Output Tokens (1M)$8.00
Cached Input (1M)$0.5000
Pricing is verified directly against Google's developer documentation and API rate sheets.

Similar & Alternative Models

Explore other frontier models from Google and comparable reasoning engines.

Browse all models
Google

The latest evolution in the Gemini 3 family, delivering state-of-the-art software engineering (73.7% DeepSWE) and agentic enterprise knowledge workflows.

1.049M ctx$0.75 / 1M tok ($1.50 reg)
Google

Universal omnimodal generation model that accepts any combination of text, audio, image, and video to generate any combination of outputs.

1.049M ctx$0.60 / 1M tok (Universal Omni)