Which is better, Claude 3.5 Sonnet or DeepSeek-V3?

Across 56 shared benchmarks tracked on Vector Wire, Claude 3.5 Sonnet leads on 18 and DeepSeek-V3 leads on 38 — DeepSeek-V3 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude 3.5 Sonnet vs DeepSeek-V3

AnthropicvsDeepSeek56 shared benchmarks18–38 head-to-head

Benchmark	Claude 3.5 Sonnet	DeepSeek-V3
AA Intelligence	10	15.4
aider_edit_acc	84.2	79.7
aider_polyglot	45.3	55.1
AIME 2024	16	59.4
AIME 2025	7.4	51.3
AIR-Bench 2024	85.9	40.8
AlpacaEval2.0 (LC-winrate)	52	70
anthropic_red_team	99.8	97.1
Arena Hard	87.6	91.4
Arena-Hard (GPT-4-1106 judge)	85.2	85.5
ArenaHard (GPT-4-1106)	85.2	85.5
Artificial Analysis Coding Index	30.2	23
bbh	93.1	87.5
bbq	94.9	96.7
C-Eval	76.7	90.1
Chinese SimpleQA (C-SimpleQA)	56.8	68
CLUEWSC	85.4	90.9
CNMO 2024	13.1	74.7
Codeforces	717	1134
Codeforces (Percentile)	20.3	58.7
Codeforces (Rating)	717	1134
DROP	87.1	91.6
DROP (3-shot F1)	88.3	91.6
DROP (F1)	88.8	91
FRAMES (Acc.)	72.5	73.3
GPQA Diamond	67.2	68.4
GSM8K	96.9	96.7
harmbench	98.1	49.7
HLE	3.9	5.2
hmmt_feb_2025	1.7	29.2
humaneval	93.7	92.1
HumanEval-Mul (Pass@1)	81.7	82.6
ifeval	90.1	87.3
IFEval (avg)	90.1	87.3
LiveCodeBench	32.8	49.6
LiveCodeBench v6	37.2	46.9
livecodebench_pass1cot	36.3	40.5
LongBench v2 overall (w/o CoT)	41	48.7
longbench_v2	41	48.7
MATH	81.3	91.2
math_500_em	78.3	94
MBPP+ (EvalPlus-augmented)	75.1	78.8
mgsm	91.6	79.8
mmlu	88.3	89.4
mmlu_redux	88.9	90.5
MMLU-Pro	78	81.2
narrativeqa	74.6	79.6
naturalquestions_closedbook	50.2	46.7
OpenBookQA	97.2	95.4
scicode	36.6	35.8
simple_safety_tests	100	95.3
simplebench	41.4	27.2
simpleqa	28.4	27.7
supergpqa	48.2	53.7
SWE-bench Verified	50.8	42
xstest	95.6	97.1

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.