Which is better, Claude Sonnet 4 or DeepSeek-V3.2?

Across 46 shared benchmarks tracked on Vector Wire, Claude Sonnet 4 leads on 15 and DeepSeek-V3.2 leads on 31 — DeepSeek-V3.2 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-V3.2

AnthropicvsDeepSeek46 shared benchmarks15–31 head-to-head

Benchmark	Claude Sonnet 4	DeepSeek-V3.2
AA Agentic Index	39.2	39.8
AA Intelligence	29.8	32.8
AA-LCR	69.7	70.7
AA-Omniscience	0.2	-22.5
AIME 2025	74	93.1
arc_agi_1	63.7	57
ARC-AGI-2	13.6	4
ArtifactsBench	57.3	55.8
Artificial Analysis Coding Index	37.6	44.2
browsecomp	12.2	51.4
browsecomp_zh	29.1	65
critpt	1.1	2.9
FinSearchComp-global	42	26.2
frontiermath_tier_4	0	2.1
GAIA (text only)	68.3	63.5
gdpval	31.2	18.8
GPQA Diamond	78	84
HLE	10.7	40.8
HLE (with tools)	20.3	40.8
HMMT 2025	15.9	90.2
IFBench	55	61
LiveCodeBench	68.5	86
LiveCodeBench v6	48.5	83.3
mmlu_redux	93.6	93.7
MMLU-Pro	84	86
Multi-SWE-Bench	35.7	37.4
OmniScience Accuracy	22.7	33
OmniScience Non-Hallucination	70.9	17.3
scicode	40	39
swe_bench_bash	64.9	60
SWE-bench Multilingual	56.9	70.2
SWE-bench Pro	42.7	15.6
SWE-bench Verified	80.2	73.1
TauBench V3 - Banking	16.7	18.8
Terminal-Bench	36.4	37.7
Terminal-Bench 2.1	36.3	46.8
Terminal-Bench Hard	31.1	35.6
theagentcompany	37	34
vectara_answer_rate	98.6	92.6
vectara_avg_summary_length	145.8	62
vectara_factual_consistency	89.7	93.7
vectara_hallucination_rate ↓	10.3	6.3
xbench_deepsearch	64.6	71
τ²-Bench	65	91
τ²-Bench Telecom (AA run)	64.6	90.6
τ³-Bench	13.8	69.2

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.