Models / DeepSeek V3.2

DeepSeek

DeepSeek V3.2

44 published results from 5 sources. Each card shows where the number comes from and what it does not measure. The overall leaderboard combines them; here each stands alone.

Provider
DeepSeek
Sources
5
Our benchmarks
0
Price per million tokens
$0.28 in · $0.42 out
OpenRouter list price, 1 Oct 2026 · 163,840-token context

Reported by others

Reported by Arena (formerly LMArena)
1,437
Overall · rank 78 of 177
Unit
Arena rating, higher is better
Range
1,434 to 1,440
Sample
47043 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not an error rate: claims that cannot be checked on the web are skipped, and preference still carries most of the weight.
Reported by Arena (formerly LMArena)
1,433
Overall · rank 84 of 177
Unit
Arena rating, higher is better
Range
1,429 to 1,436
Sample
41092 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not an error rate: claims that cannot be checked on the web are skipped, and preference still carries most of the weight.
Reported by Arena (formerly LMArena)
1,422
Business, management and finance · rank 100 of 402
Unit
Arena rating, higher is better
Range
1,415 to 1,428
Sample
9053 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,423
Business, management and finance · rank 99 of 402
Unit
Arena rating, higher is better
Range
1,416 to 1,430
Sample
8104 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,400
Creative writing · rank 79 of 407
Unit
Arena rating, higher is better
Range
1,393 to 1,408
Sample
6859 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,391
Creative writing · rank 94 of 407
Unit
Arena rating, higher is better
Range
1,383 to 1,399
Sample
6383 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,448
Expert prompts · rank 84 of 359
Unit
Arena rating, higher is better
Range
1,438 to 1,458
Sample
3419 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,453
Expert prompts · rank 77 of 359
Unit
Arena rating, higher is better
Range
1,442 to 1,464
Sample
2880 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,420
Instruction following · rank 86 of 409
Unit
Arena rating, higher is better
Range
1,414 to 1,426
Sample
13530 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,420
Instruction following · rank 87 of 409
Unit
Arena rating, higher is better
Range
1,414 to 1,425
Sample
11720 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,425
Overall · rank 105 of 409
Unit
Arena rating, higher is better
Range
1,421 to 1,428
Sample
48010 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,423
Overall · rank 109 of 409
Unit
Arena rating, higher is better
Range
1,419 to 1,426
Sample
41980 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,410
Writing, literature and language · rank 86 of 408
Unit
Arena rating, higher is better
Range
1,403 to 1,416
Sample
10655 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Arena (formerly LMArena)
1,403
Writing, literature and language · rank 91 of 408
Unit
Arena rating, higher is better
Range
1,397 to 1,410
Sample
9719 votes
Configuration
DeepSeek V3.2
Measured
25 Sep 2026
Not shown
Not accuracy or correctness: it ranks which answer voters preferred, with answer length and formatting controlled for.
Reported by Artificial Analysis
44.2
Artificial Analysis Coding Index · rank 105 of 151
Unit
index score, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not your codebase or tools.
Reported by Artificial Analysis
16.0
Artificial Analysis Intelligence Index · rank 208 of 314
Unit
index score, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not business work, and a blend: read the parts for any one task.
Reported by Artificial Analysis
21.5
Artificial Analysis Intelligence Index · rank 166 of 314
Unit
index score, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not business work, and a blend: read the parts for any one task.
Reported by Artificial Analysis
75.1%
GPQA Diamond · rank 188 of 276
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not applied work; multiple-choice science questions.
Reported by Artificial Analysis
84.0%
GPQA Diamond · rank 126 of 276
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not applied work; multiple-choice science questions.
Reported by Artificial Analysis
11.2%
Humanity's Last Exam · rank 219 of 314
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not everyday work; academic questions at the edge of expertise.
Reported by Artificial Analysis
24.6%
Humanity's Last Exam · rank 150 of 314
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not everyday work; academic questions at the edge of expertise.
Reported by Artificial Analysis
49.0%
IFBench · rank 121 of 215
Unit
% of instructions, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not judgement about what an instruction meant.
Reported by Artificial Analysis
60.7%
IFBench · rank 84 of 215
Unit
% of instructions, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not judgement about what an instruction meant.
Reported by Artificial Analysis
45.7%
Long-context reasoning (AA-LCR) · rank 241 of 309
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not retrieval over your own document store.
Reported by Artificial Analysis
73.3%
Long-context reasoning (AA-LCR) · rank 145 of 309
Unit
% of questions, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not retrieval over your own document store.
Reported by Artificial Analysis
46.8%
Terminal-Bench 2.1 · rank 104 of 151
Unit
% of tasks, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not other harnesses or tools; superseded by 4.0 for newer models.
Reported by Artificial Analysis
32.6%
Terminal-Bench Hard · rank 85 of 213
Unit
% of tasks, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not other harnesses; Artificial Analysis no longer runs it on new models.
Reported by Artificial Analysis
35.6%
Terminal-Bench Hard · rank 66 of 213
Unit
% of tasks, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not other harnesses; Artificial Analysis no longer runs it on new models.
Reported by Artificial Analysis
78.9%
Τ²-bench telecom · rank 93 of 214
Unit
% of tasks, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
1 Oct 2026
Not shown
Not your policies or systems; no longer run on new models.
Reported by Artificial Analysis
90.6%
Τ²-bench telecom · rank 50 of 214
Unit
% of tasks, higher is better
Configuration
DeepSeek V3.2
Measured
1 Oct 2026
Not shown
Not your policies or systems; no longer run on new models.
Reported by OpenHands Index
45.7%
Average · rank 21 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by OpenHands Index
27.9%
Front end (SWE-Bench Multimodal) · rank 24 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by OpenHands Index
25.0%
Greenfield (Commit0) · rank 14 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by OpenHands Index
50.3%
Information gathering (GAIA) · rank 22 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by OpenHands Index
71.6%
Issue resolution (SWE-Bench) · rank 24 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by OpenHands Index
53.6%
Testing (SWT-Bench) · rank 25 of 34
Unit
% resolved, higher is better
Configuration
DeepSeek V3.2 (thinking reasoning, openhands)
Measured
30 Jun 2026
Not shown
Not the model alone: the score is for the model inside the OpenHands agent, at the SDK version shown.
Reported by UGI Leaderboard
26.0%
Requested-length error · rank 233 of 370
Unit
% off the requested word count, lower is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
3 Dec 2025
Not shown
Not other format limits such as character counts or bullet counts.
Reported by UGI Leaderboard
29.0%
Requested-length error · rank 248 of 370
Unit
% off the requested word count, lower is better
Configuration
DeepSeek V3.2
Measured
3 Dec 2025
Not shown
Not other format limits such as character counts or bullet counts.
Reported by UGI Leaderboard
0.34
Style adherence · rank 232 of 370
Unit
score from 0 to 1, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
3 Dec 2025
Not shown
Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
Reported by UGI Leaderboard
0.36
Style adherence · rank 111 of 370
Unit
score from 0 to 1, higher is better
Configuration
DeepSeek V3.2
Measured
3 Dec 2025
Not shown
Not brand voice on your own examples: UGI's prompts are private and lean towards creative writing.
Reported by UGI Leaderboard
56.3
Writing score · rank 124 of 370
Unit
score out of 100, higher is better
Configuration
DeepSeek V3.2 (no reasoning)
Measured
3 Dec 2025
Not shown
Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
Reported by UGI Leaderboard
57.3
Writing score · rank 117 of 370
Unit
score out of 100, higher is better
Configuration
DeepSeek V3.2
Measured
3 Dec 2025
Not shown
Not business copy quality: UGI's prompts are private and lean towards creative writing, and models that often refuse get no score.
Reported by Vectara Hallucination Leaderboard
92.6%
Answer rate · rank 98 of 108
Unit
% of documents, higher is better
Configuration
DeepSeek V3.2
Measured
22 Sep 2026
Not shown
Not a quality score: a low rate usually means content filters were triggered, and hallucination rates are measured on answered documents only.
Reported by Vectara Hallucination Leaderboard
6.3%
Hallucination rate · rank 27 of 108
Unit
% of summaries, lower is better
Configuration
DeepSeek V3.2
Measured
22 Sep 2026
Not shown
Not errors in open questions or other tasks: only summarisation, judged by Vectara's own model (HHEM-2.3), not by people, on news-style documents rather than your data.

Compare DeepSeek V3.2 with