Which is better, Claude Sonnet 4.6 or DeepSeek-V4-Pro?

Across 54 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.6 leads on 30 and DeepSeek-V4-Pro leads on 23 — Claude Sonnet 4.6 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 4.6 vs DeepSeek-V4-Pro

AnthropicvsDeepSeek54 shared benchmarks30–23 head-to-head

Benchmark	Claude Sonnet 4.6	DeepSeek-V4-Pro
AA Agentic Index	61.6	63.3
AA Intelligence	48.4	53
AA-LCR	74	70
AA-Omniscience	12.2	-10.6
AgentWorldBench - Android	58	55.2
AgentWorldBench - MCP	70	63.3
AgentWorldBench - OS	63.2	63.7
AgentWorldBench - Overall	56	53
AgentWorldBench - Search	28.8	27.6
AgentWorldBench - SWE	64.5	59.4
AgentWorldBench - Terminal	57	51.3
AgentWorldBench - Web	50.8	50.3
apexAgents	28	24.3
arena_elo	1472	1457
arena_text_factuality	1460	1450
Artificial Analysis Coding Index	63	59.4
browsecomp	76.2	83.4
coding_arena_elo	1523	1582
critpt	3.1	13
gdpval	54.8	49
GDPval-AA (Elo)	1676	1554
GDPval-AA v2	1381	1307
GPQA Diamond	89.9	90.5
HLE	49	48.2
HLE (with tools)	46.8	48.2
IFBench	56.6	76.5
itbenchSre	39.8	38.3
livebench	75.5	73.6
livebench_agentic_coding	42.6	42.6
livebench_coding	79.3	70
livebench_data_analysis	78	74.5
livebench_instruction_following	63.2	62.4
livebench_language	76.1	78.1
livebench_math	87	90.7
livebench_reasoning	84.8	82.7
MCP Atlas	69.5	74.2
OmniScience Accuracy	40.9	43
OmniScience Non-Hallucination	51.6	12.2
scicode	47	50
simpleqa_verified	29	57.9
SWE-bench Pro	58.1	55.4
SWE-bench Verified	79.6	80.6
TauBench V3 - Banking	34.4	30.1
Terminal-Bench 2.0	59.1	67.9
Terminal-Bench 2.1	71.2	72.1
Terminal-Bench Hard	59.1	46.2
toolathlon	49.4	51.8
usamo_2026	55	60.7
vectara_answer_rate	99.9	97.2
vectara_avg_summary_length	114.7	153.8
vectara_factual_consistency	89.4	91.4
vectara_hallucination_rate ↓	10.6	8.6
τ²-Bench Telecom (AA run)	97.9	96.2
τ³-Bench	30.5	25.8

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.