Compare · ModelsLive · 3 picked · head to head
DeepSeek V3.2 Exp vs Gemini 2.5 Flash vs Claude Sonnet 4
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
Gemini 2.5 Flash wins on 6/15 benchmarks
Gemini 2.5 Flash wins 6 of 15 shared benchmarks. Leads in knowledge · math.
Category leads
coding·DeepSeek V3.2 Expknowledge·Gemini 2.5 Flashagentic·DeepSeek V3.2 Expreasoning·Claude Sonnet 4arena·DeepSeek V3.2 Expmath·Gemini 2.5 Flash
Hype vs Reality
Attention vs performance
DeepSeek V3.2 Exp
#80 by perf·no signal
Gemini 2.5 Flash
#144 by perf·#14 by attention
Claude Sonnet 4
#117 by perf·no signal
Best value
DeepSeek V3.2 Exp
5.5x better value than Gemini 2.5 Flash
DeepSeek V3.2 Exp
156.5 pts/$
$0.34/M
Gemini 2.5 Flash
28.6 pts/$
$1.40/M
Claude Sonnet 4
5.0 pts/$
$9.00/M
Vendor risk
Mixed exposure
One or more vendors flagged
DeepSeek
$3.4B·Tier 1
Google DeepMind
$4.00T·Tier 1
Anthropic
$380.0B·Tier 1
Head to head
15 benchmarks · 3 models
DeepSeek V3.2 ExpGemini 2.5 FlashClaude Sonnet 4
Aider polyglot
DeepSeek V3.2 Exp leads by +12.9
Aider Polyglot · measures how well AI models can edit code across multiple programming languages using the Aider coding assistant framework.
DeepSeek V3.2 Exp
74.2
Gemini 2.5 Flash
47.1
Claude Sonnet 4
61.3
Fiction.LiveBench
DeepSeek V3.2 Exp leads by +36.1
Fiction.LiveBench · a continuously updated benchmark using recently published fiction to test reading comprehension and reasoning, preventing data contamination.
DeepSeek V3.2 Exp
83.3
Gemini 2.5 Flash
47.2
Claude Sonnet 4
46.9
The Agent Company
DeepSeek V3.2 Exp leads by +1.8
The Agent Company · tests AI agents on realistic corporate tasks like email management, code review, data analysis, and cross-tool workflows.
DeepSeek V3.2 Exp
42.9
Gemini 2.5 Flash
41.1
Claude Sonnet 4
33.1
WeirdML
Claude Sonnet 4 leads by +5.2
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
DeepSeek V3.2 Exp
39.5
Gemini 2.5 Flash
41.0
Claude Sonnet 4
46.1
ARC-AGI
Claude Sonnet 4 leads by +7.7
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
Gemini 2.5 Flash
32.3
Claude Sonnet 4
40.0
ARC-AGI-2
Claude Sonnet 4 leads by +3.4
ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data.
Gemini 2.5 Flash
2.5
Claude Sonnet 4
5.9
Chatbot Arena Elo · Overall
DeepSeek V3.2 Exp leads by +11.8
DeepSeek V3.2 Exp
1422.8
Gemini 2.5 Flash
1411.0
DeepResearch Bench
Claude Sonnet 4 leads by +18.6
DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses.
Gemini 2.5 Flash
29.2
Claude Sonnet 4
47.8
FrontierMath-2025-02-28-Private
Gemini 2.5 Flash leads by +0.7
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
Gemini 2.5 Flash
4.8
Claude Sonnet 4
4.1
FrontierMath-Tier-4-2025-07-01-Private
Gemini 2.5 Flash leads by +4.1
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
Gemini 2.5 Flash
4.2
Claude Sonnet 4
0.1
GeoBench
Gemini 2.5 Flash leads by +36.0
GeoBench · tests geographic knowledge and spatial reasoning across countries, landmarks, coordinates, and geopolitical understanding.
Gemini 2.5 Flash
73.0
Claude Sonnet 4
37.0
HLE
Gemini 2.5 Flash leads by +4.5
HLE (Humanity's Last Exam) · a reasoning benchmark designed to be the hardest public evaluation of AI. Questions span mathematics, physics, philosophy, and logic · curated to be at or beyond the frontier of human expert capability. Tested with and without tool augmentation. Claude Opus 4.7 scores 46.9% without tools and 54.7% with tools · making it one of the few benchmarks where the top score is below 60%.
Gemini 2.5 Flash
7.7
Claude Sonnet 4
3.1
OTIS Mock AIME 2024-2025
Gemini 2.5 Flash leads by +2.0
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
Gemini 2.5 Flash
73.0
Claude Sonnet 4
71.1
SimpleBench
Claude Sonnet 4 leads by +5.2
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
Gemini 2.5 Flash
29.4
Claude Sonnet 4
34.6
VPCT
Gemini 2.5 Flash leads by +6.0
VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations.
Gemini 2.5 Flash
7.0
Claude Sonnet 4
1.0
Full benchmark table
| Benchmark | DeepSeek V3.2 Exp | Gemini 2.5 Flash | Claude Sonnet 4 |
|---|---|---|---|
Aider polyglot Aider Polyglot · measures how well AI models can edit code across multiple programming languages using the Aider coding assistant framework. | 74.2 | 47.1 | 61.3 |
Fiction.LiveBench Fiction.LiveBench · a continuously updated benchmark using recently published fiction to test reading comprehension and reasoning, preventing data contamination. | 83.3 | 47.2 | 46.9 |
The Agent Company The Agent Company · tests AI agents on realistic corporate tasks like email management, code review, data analysis, and cross-tool workflows. | 42.9 | 41.1 | 33.1 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 39.5 | 41.0 | 46.1 |
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | — | 32.3 | 40.0 |
ARC-AGI-2 ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data. | — | 2.5 | 5.9 |
Chatbot Arena Elo · Overall | 1422.8 | 1411.0 | — |
DeepResearch Bench DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses. | — | 29.2 | 47.8 |
FrontierMath-2025-02-28-Private FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning. | — | 4.8 | 4.1 |
FrontierMath-Tier-4-2025-07-01-Private FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning. | — | 4.2 | 0.1 |
GeoBench GeoBench · tests geographic knowledge and spatial reasoning across countries, landmarks, coordinates, and geopolitical understanding. | — | 73.0 | 37.0 |
HLE HLE (Humanity's Last Exam) · a reasoning benchmark designed to be the hardest public evaluation of AI. Questions span mathematics, physics, philosophy, and logic · curated to be at or beyond the frontier of human expert capability. Tested with and without tool augmentation. Claude Opus 4.7 scores 46.9% without tools and 54.7% with tools · making it one of the few benchmarks where the top score is below 60%. | — | 7.7 | 3.1 |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | — | 73.0 | 71.1 |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | — | 29.4 | 34.6 |
VPCT VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations. | — | 7.0 | 1.0 |
Pricing · per 1M tokens · projected $/mo at 10M tokens
| Model | Input | Output | Context | Projected $/mo |
|---|---|---|---|---|
| $0.27 | $0.41 | 164K tokens (~82 books) | $3.05 | |
| $0.30 | $2.50 | 1.0M tokens (~524 books) | $8.50 | |
| $3.00 | $15.00 | 1.0M tokens (~500 books) | $60.00 |