Which is better, Claude Sonnet 4 or DeepSeek-V3?

Across 59 shared benchmarks tracked on Vector Wire, Claude Sonnet 4 leads on 48 and DeepSeek-V3 leads on 10 — Claude Sonnet 4 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude Sonnet 4 vs DeepSeek-V3

AnthropicvsDeepSeek59 shared benchmarks48–10 head-to-head

Benchmark	Claude Sonnet 4	DeepSeek-V3
AA Agentic Index	39.2	1.6
AA Intelligence	29.8	15.4
AA-LCR	69.7	41.3
AA-Omniscience	0.2	-37.6
AceBench	76.2	72.7
aider_polyglot	70.7	55.1
AIME 2024	43.4	59.4
AIME 2025	74	51.3
AIR-Bench 2024	88.3	40.8
anthropic_red_team	99.2	97.1
Artificial Analysis Coding Index	37.6	23
AutoLogi	89.8	88.9
bbq	97.9	96.7
CNMO 2024	60.4	74.7
critpt	1.1	0
gdpval	31.2	0
GPQA Diamond	78	68.4
harmbench	98	49.7
HLE	10.7	5.2
HMMT 2025	15.9	27.5
IFBench	55	41
ifeval	87.6	87.3
livebench	74.8	72.4
LiveCodeBench	68.5	49.6
LiveCodeBench v6	48.5	46.9
livecodebench_easy	98.8	83.3
livecodebench_hard	32.6	14.2
livecodebench_medium	76	53.5
math_500_em	94	94
mmlu	91.5	89.4
mmlu_redux	93.6	90.5
MMLU-Pro	84	81.2
mmmlu	86.5	79.4
multichallenge	57.1	31.4
MultiPL-E	88.6	83.1
OJBench	15.3	24
OmniScience Accuracy	22.7	25.4
OmniScience Non-Hallucination	70.9	23.3
PolyMath-en	52.8	59.5
scicode	40	35.8
simple_safety_tests	100	95.3
simplebench	45.5	27.2
simpleqa	15.9	27.7
supergpqa	55.7	53.7
SWE-bench Verified	80.2	42
Tau2 airline	55.5	39
Tau2 retail	75	69.1
TauBench V3 - Banking	16.7	4.7
Terminal-Bench 2.1	36.3	16.9
Terminal-Bench Hard	31.1	15.2
vectara_answer_rate	98.6	97.5
vectara_avg_summary_length	145.8	81.7
vectara_factual_consistency	89.7	93.9
vectara_hallucination_rate ↓	10.3	6.1
xstest	97.2	97.1
ZebraLogic	73.7	84
τ²-Bench	65	32.5
τ²-Bench Telecom (AA run)	64.6	47.1
τ³-Bench	13.8	4.7

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.