Which is better, Claude Sonnet 4 or DeepSeek-R1?

Across 57 shared benchmarks tracked on Vector Wire, Claude Sonnet 4 leads on 36 and DeepSeek-R1 leads on 21 — 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-R1

AnthropicvsDeepSeek57 shared benchmarks36–21 head-to-head

Benchmark	Claude Sonnet 4	DeepSeek-R1
AA Agentic Index	39.2	3.1
AA Intelligence	29.8	18.6
AA-LCR	69.7	56
AA-Omniscience	0.2	-31.3
aider_polyglot	70.7	71.6
AIME 2024	43.4	91.4
AIME 2025	74	87.5
AIR-Bench 2024	88.3	52.9
anthropic_red_team	99.2	99.1
arc_agi_1	63.7	21.2
ARC-AGI-2	13.6	1.3
Artificial Analysis Coding Index	37.6	24.6
bbq	97.9	96.6
browsecomp	12.2	8.9
browsecomp_zh	29.1	35.7
CNMO 2024	60.4	78.8
critpt	1.1	0.6
Fortress	24.4	74.4
gdpval	31.2	1.5
GPQA Diamond	78	81
harmbench	98	54.6
HLE	10.7	17.7
HMMT 2025	15.9	79.4
IFBench	55	39
ifeval	87.6	83.3
LiveCodeBench	68.5	84.4
livecodebench_easy	98.8	99.2
livecodebench_hard	32.6	63.6
livecodebench_medium	76	90.9
math_500_em	94	98
mmlu	91.5	92.9
mmlu_redux	93.6	93.4
MMLU-Pro	84	85
multichallenge	57.1	45
MultiNRC	18.4	27.6
OmniScience Accuracy	22.7	30.7
OmniScience Non-Hallucination	70.9	10.5
scicode	40	35.7
simple_safety_tests	100	98.3
simplebench	45.5	40.8
simpleqa	15.9	92.3
SWE-bench Multilingual	56.9	30.5
SWE-bench Verified	80.2	57.6
TAU-bench (airline)	60	53.5
TAU-bench (retail)	80.5	63.9
TauBench V3 - Banking	16.7	6.4
Terminal-Bench	36.4	5.7
Terminal-Bench 2.1	36.3	19.1
Terminal-Bench Hard	31.1	6.1
vectara_answer_rate	98.6	97
vectara_avg_summary_length	145.8	93.5
vectara_factual_consistency	89.7	88.7
vectara_hallucination_rate ↓	10.3	11.3
xstest	97.2	98.8
ZebraLogic	73.7	95.1
τ²-Bench Telecom (AA run)	64.6	11.4
τ³-Bench	13.8	6.4

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.