Inference
EN

DeepSeek V3.1 vs DeepSeek V3.1 (Reasoning)

Change models
DeepSeek V3.1DeepSeek
DeepSeek V3.1 (Reasoning)DeepSeek
IndexAll categories, out of 100.
37.9 ±5.8
43.0 ±6.4 best
GapIndex points below the top.
−5.1 · no clear gap
Top
CoverageIndex weight that has results.
85%
85%
SpeedOutput tokens per second.
98/s
—
EloLMArena rating and rank.
1420
—
ReleasedMonth of release.
2025-08
2025-08
Input / 1MUS dollars per 1M input tokens.
—
—
Output / 1MUS dollars per 1M output tokens.
—
—
1K chats2K tokens in, 500 out, each.
—
—
1K documents20K tokens in, 1K out, each.
—
—
1K agent tasks200K tokens in, 20K out, each.
—
—
ContextMaximum tokens in one request.
128K
128K
AgenticMulti-step tasks with tools.
38.0 best
27.6
CodingCode writing and repair.
43.0
51.4 best
ReasoningLogic problems and puzzles.
38.2
43.5 best
MultimodalTasks with images and text.
—
—
KnowledgeFacts and expert knowledge.
41.5
47.0 best
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
24.3
27.8 best
MathMath problems.
29.9
46.5 best

Benchmarks

12 shared
Only DeepSeek V3.1: 5Only DeepSeek V3.1 (Reasoning): 3
DeepSeek V3.1
DeepSeek V3.1 (Reasoning)
Artificial Analysis Humanity's Last ExamKnowledge
6.7%165 / 222
14.3% best128 / 222
Artificial Analysis Long Context ReasoningReasoning
47.0%161 / 217
56.7% best144 / 217
Critical Physics TasksReasoning
0.0%150 / 214
2.0% best105 / 214
Artificial Analysis GPQA DiamondKnowledge
73.5%124 / 199
77.9% best106 / 199
Artificial Analysis Omniscience AccuracyKnowledge
23.1%115 / 164
29.0% best82 / 164
τ²-Bench Tool-Agent-User EvaluationAgentic
34.8%109 / 149
37.4% best105 / 149
Artificial Analysis IFBenchInstruction
37.8%115 / 146
41.5% best98 / 146
Graduate-Level Google-Proof Q&AKnowledge
74.9%82 / 141
80.1% best66 / 141
Massive Multitask Language Understanding ProfessionalKnowledge
83.7%28 / 75
84.8% best23 / 75
American Invitational Mathematics Examination 2025Math
49.8%42 / 53
88.4% best20 / 53
American Invitational Mathematics Examination 2024Math
66.3%20 / 27
93.1% best3 / 27
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for CodeCoding
56.4%8 / 12
74.8% best2 / 12