Which is better, Claude Sonnet 4.5 or DeepSeek-V3.2?

Across 83 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 45 and DeepSeek-V3.2 leads on 37 — Claude Sonnet 4.5 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 4.5 vs DeepSeek-V3.2

AnthropicvsDeepSeek83 shared benchmarks45–37 head-to-head

Benchmark	Claude Sonnet 4.5	DeepSeek-V3.2
AA Agentic Index	50.6	39.8
AA Intelligence	37.4	32.8
AA-LCR	68.3	70.7
AA-Omniscience	-0.1	-22.5
AIME 2025	88	93.1
AIME25 no tools	88	89.3
AIME25 w/ python	100	58.1
arc_agi_1	63.7	57
ARC-AGI-2	13.6	4
arena_text_factuality	1448	1433
Arena-Hard (Creative Writing)	76.7	88.8
Arena-Hard (Hard Prompt)	63.3	53.4
ArtifactsBench	61.5	55.8
Artificial Analysis Coding Index	52.1	44.2
browsecomp	24.1	51.4
BrowseComp (context management)	26.1	67.6
BrowseComp w/ tools	24.1	40.1
browsecomp_zh	42.4	65
BrowseComp-ZH w/ tools	42.4	47.9
coding_arena_elo	1392	1368
critpt	1.1	2.9
FinSearchComp-global	60.8	26.2
FinSearchComp-T3	44	27
FinSearchComp-T3 w/ tools	44	27
Frames	85	80.2
Frames w/ tools	85	80.2
frontiermath_tier_4	4.2	2.1
GAIA (text only)	71.2	63.5
gdpval	40.4	18.8
GPQA (unspecified)	83.4	79.9
GPQA Diamond	83.4	84
HealthBench	44.2	46.9
HealthBench no tools	44.2	46.9
HLE	19.8	40.8
HLE (with tools)	32	40.8
HMMT 2025	74.6	90.2
hmmt_feb_2025	79.2	92.5
hmmt_nov_2025	81.7	90.2
HMMT25 no tools	74.6	83.6
HMMT25 w/ python	88.8	49.5
IFBench	57.3	61
imo_answer_bench	65.9	78.3
LiveCodeBench	71	86
LiveCodeBench v6	64	83.3
LiveCodeBenchV6 no tools	64	74.1
longbench_v2	61.8	59.8
Longform Writing eval (Kimi K2 Thinking system card)	79.8	72.5
MCP Atlas	59.5	62.2
mmlu_redux	95.6	93.7
MMLU-Pro	88.2	86
mrcr	55.4	55.5
Multi-SWE-Bench	44.3	37.4
OctoCodingbench	22.8	26
OJ-Bench (cpp)	30.4	54.7
OJ-Bench (cpp) no tools	30.4	38.2
OmniScience Accuracy	32.9	33
OmniScience Non-Hallucination	50.9	17.3
scicode	45	39
Seal-0	53.4	49.5
Seal-0 w/ tools	53.4	38.5
simpleqa_verified	23.6	27.5
swe_bench_bash	71.4	60
SWE-bench Multilingual	68	70.2
SWE-bench Pro	43.6	15.6
SWE-bench Verified	82	73.1
SWE-Perf	3	0.9
SWT-bench	69.5	62
TauBench V3 - Banking	24.5	18.8
Terminal-Bench	51	37.7
Terminal-Bench 2.0	50	46.4
Terminal-Bench 2.1	55.8	46.8
Terminal-Bench Hard	35.6	35.6
Terminal-Bench w/ simulated tools (JSON)	51	37.7
theagentcompany	41	34
toolathlon	41	35.2
vectara_answer_rate	95.6	92.6
vectara_avg_summary_length	127.8	62
vectara_factual_consistency	88	93.7
vectara_hallucination_rate ↓	12	6.3
xbench_deepsearch	66	71
τ²-Bench	98	91
τ²-Bench Telecom (AA run)	78.1	90.6
τ³-Bench	19	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.