Which is better, DeepSeek-V3.2 or GPT-5.4?

Across 46 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 3 and GPT-5.4 leads on 43 — GPT-5.4 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 GPT-5.4

DeepSeekvsOpenAI46 shared benchmarks3–43 head-to-head

Benchmark	DeepSeek-V3.2	GPT-5.4
AA Agentic Index	39.8	58.2
AA Intelligence	32.8	53.1
AA-LCR	70.7	77.7
AA-Omniscience	-22.5	5.8
aime_2026	95.1	99.2
apexAgents	14.5	33.3
arc_agi_1	57	93.7
ARC-AGI-2	4	74
arena_text_factuality	1433	1476
Artificial Analysis Coding Index	44.2	71.1
browsecomp	51.4	82.7
browsecomp_with_context_manager	67.6	82.7
coding_arena_elo	1368	1457
critpt	2.9	23.4
cybergym	17.3	66.3
deepsearchqa_f1	60.9	78.6
frontiermath_tier_4	2.1	27.1
gdpval	18.8	50.1
GPQA Diamond	84	93
HLE	40.8	43.7
HLE (with tools)	40.8	52.1
hmmt_feb_2026	84.1	97.7
hmmt_nov_2025	90.2	95.8
IFBench	61	73.9
imo_answer_bench	78.3	91.4
MCP Atlas	62.2	70.6
mcpmark	38	62.5
MMLU-Pro	86	87.5
OmniScience Accuracy	33	50.9
OmniScience Non-Hallucination	17.3	17.4
scicode	39	56.6
simpleqa_verified	27.5	45.3
SWE-bench Multilingual	70.2	71.7
SWE-bench Pro	15.6	59.1
TauBench V3 - Banking	18.8	39.6
Terminal-Bench 2.0	46.4	75.1
Terminal-Bench 2.1	46.8	78.3
Terminal-Bench Hard	35.6	57.6
tool_decathlon	35.2	54.6
toolathlon	35.2	54.6
vectara_answer_rate	92.6	99.9
vectara_avg_summary_length	62	81.7
vectara_factual_consistency	93.7	93
vectara_hallucination_rate ↓	6.3	7
τ²-Bench Telecom (AA run)	90.6	87.1
τ³-Bench	69.2	72.9

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.