Which is better, DeepSeek-V3.2 or DeepSeek-V4-Pro?

Across 42 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 5 and DeepSeek-V4-Pro leads on 37 — DeepSeek-V4-Pro leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V3.2 vs DeepSeek-V4-Pro

DeepSeekvsDeepSeek42 shared benchmarks5–37 head-to-head

Benchmark	DeepSeek-V3.2	DeepSeek-V4-Pro
AA Agentic Index	39.8	63.3
AA Intelligence	32.8	53
AA-LCR	70.7	70
AA-Omniscience	-22.5	-10.6
aime_2026	95.1	96.7
apexAgents	14.5	24.3
arena_text_factuality	1433	1450
Artificial Analysis Coding Index	44.2	59.4
browsecomp	51.4	83.4
coding_arena_elo	1368	1582
critpt	2.9	13
cybergym	17.3	52.7
gdpval	18.8	49
GPQA Diamond	84	90.5
HLE	40.8	48.2
HLE (with tools)	40.8	48.2
hmmt_feb_2026	84.1	95.2
hmmt_nov_2025	90.2	94.4
IFBench	61	76.5
imo_answer_bench	78.3	89.8
LiveCodeBench	86	93.5
MCP Atlas	62.2	74.2
MMLU-Pro	86	87.5
OmniScience Accuracy	33	43
OmniScience Non-Hallucination	17.3	12.2
scicode	39	50
simpleqa_verified	27.5	57.9
SWE-bench Multilingual	70.2	76.2
SWE-bench Pro	15.6	55.4
SWE-bench Verified	73.1	80.6
TauBench V3 - Banking	18.8	30.1
Terminal-Bench 2.0	46.4	67.9
Terminal-Bench 2.1	46.8	72.1
Terminal-Bench Hard	35.6	46.2
tool_decathlon	35.2	52.8
toolathlon	35.2	51.8
vectara_answer_rate	92.6	97.2
vectara_avg_summary_length	62	153.8
vectara_factual_consistency	93.7	91.4
vectara_hallucination_rate ↓	6.3	8.6
τ²-Bench Telecom (AA run)	90.6	96.2
τ³-Bench	69.2	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.