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

Across 63 shared benchmarks tracked on Vector Wire, Claude Opus 4.5 leads on 51 and DeepSeek-V3.2 leads on 11 — Claude Opus 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 Opus 4.5 vs DeepSeek-V3.2

AnthropicvsDeepSeek63 shared benchmarks51–11 head-to-head

Benchmark	Claude Opus 4.5	DeepSeek-V3.2
AA Agentic Index	59.6	39.8
AA Intelligence	41.9	32.8
AA-LCR	76	70.7
AA-Omniscience	14	-22.5
AIME 2025	92.8	93.1
AIME 2026 I	93.3	92.7
aime_2026	95.1	95.1
AIME25 no tools	91	89.3
arc_agi_1	82	57
ARC-AGI-2	37.6	4
arena_text_factuality	1458	1433
Artificial Analysis Coding Index	47.8	44.2
browsecomp	37	51.4
BrowseComp (context management)	59.2	67.6
browsecomp_with_context_manager	67.8	67.6
browsecomp_zh	62.4	65
coding_arena_elo	1495	1368
critpt	4.6	2.9
cybergym	50.6	17.3
deepsearchqa_f1	76.1	60.9
FinSearchCompT2&T3	66.2	59.1
frontiermath_tier_4	4.2	2.1
gdpval	47.3	18.8
GPQA Diamond	87	84
HLE	30.8	40.8
HLE (with tools)	43.4	40.8
hmmt_feb_2025	92.9	92.5
hmmt_feb_2026	85.3	84.1
hmmt_nov_2025	93.3	90.2
IFBench	58	61
imo_answer_bench	84	78.3
LiveCodeBench	87	86
LiveCodeBench v6	84.8	83.3
longbench_v2	64.4	59.8
MCP Atlas	69.8	62.2
mmlu_redux	95.6	93.7
MMLU-Pro	90	86
Multi-SWE-Bench	50	37.4
OctoCodingbench	36.2	26
OJ-Bench (cpp)	54.6	54.7
OmniScience Accuracy	46.6	33
OmniScience Non-Hallucination	39	17.3
PaperBench	72.9	47.1
scicode	50	39
Seal-0	47.7	49.5
simpleqa_verified	41.8	27.5
swe_bench_bash	76.8	60
SWE-bench Multilingual	77.5	70.2
SWE-bench Pro	57.1	15.6
SWE-bench Verified	81.5	73.1
SWE-Perf	4.7	0.9
SWT-bench	80.2	62
Terminal-Bench 2.0	59.3	46.4
Terminal-Bench Hard	47	35.6
tool_decathlon	43.5	35.2
toolathlon	43.5	35.2
vectara_answer_rate	98.7	92.6
vectara_avg_summary_length	114.5	62
vectara_factual_consistency	89.1	93.7
vectara_hallucination_rate ↓	10.9	6.3
WideSearch (item-f1)	76.2	32.5
τ²-Bench	98.2	91
τ²-Bench Telecom (AA run)	89.5	90.6

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.