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

Across 52 shared benchmarks tracked on Vector Wire, Claude 4 Opus leads on 41 and DeepSeek-V3 leads on 11 — Claude 4 Opus leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude 4 Opus vs DeepSeek-V3

AnthropicvsDeepSeek52 shared benchmarks41–11 head-to-head

Benchmark	Claude 4 Opus	DeepSeek-V3
AA Intelligence	31.7	15.4
AA-LCR	40	41.3
AceBench	75.6	72.7
aider_polyglot	70.7	55.1
AIME 2024	76	59.4
AIME 2025	75.5	51.3
AIR-Bench 2024	85.7	40.8
anthropic_red_team	98.9	97.1
Artificial Analysis Coding Index	34	23
AutoLogi	86.1	88.9
bbq	99.3	96.7
CNMO 2024	57.6	74.7
GPQA Diamond	79.6	68.4
harmbench	91.7	49.7
HLE	12.3	5.2
HMMT 2025	15.9	27.5
hmmt_feb_2025	60	29.2
IFBench	53.7	41
ifeval	87.4	87.3
livebench	74.6	72.4
LiveCodeBench	70.4	49.6
LiveCodeBench v6	47.4	46.9
livecodebench_easy	99.1	83.3
livecodebench_hard	33.1	14.2
livecodebench_medium	80.4	53.5
longbench_v2	55.6	48.7
math_500_em	98.2	94
mmlu	92.9	89.4
mmlu_redux	94.2	90.5
MMLU-Pro	86.6	81.2
mmmlu	88.8	79.4
multichallenge	58.6	31.4
MultiPL-E	89.6	83.1
OJBench	19.6	24
PolyMath-en	49.8	59.5
scicode	40.9	35.8
simple_safety_tests	100	95.3
simplebench	58.8	27.2
simpleqa	22.8	27.7
supergpqa	56.5	53.7
SWE-bench Verified	79.4	42
Tau2 airline	60	39
Tau2 retail	81.8	69.1
Terminal-Bench Hard	31.1	15.2
vectara_answer_rate	91	97.5
vectara_avg_summary_length	123.2	81.7
vectara_factual_consistency	88	93.9
vectara_hallucination_rate ↓	12	6.1
xstest	97	97.1
ZebraLogic	95.1	84
τ²-Bench	57	32.5
τ²-Bench Telecom (AA run)	73.4	47.1

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.