Which is better, DeepSeek-V3.2 or GPT-5.2?

Across 63 shared benchmarks tracked on Vector Wire, DeepSeek-V3.2 leads on 9 and GPT-5.2 leads on 54 — GPT-5.2 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-5.2

DeepSeekvsOpenAI63 shared benchmarks9–54 head-to-head

Benchmark	DeepSeek-V3.2	GPT-5.2
AA Agentic Index	39.8	60.2
AA Intelligence	32.8	43.3
AA-LCR	70.7	79.3
AA-Omniscience	-22.5	0.3
AIME 2025	93.1	100
aime_2026	95.1	98.3
AIME25 no tools	89.3	98
apexAgents	14.5	23
arc_agi_1	57	86.2
ARC-AGI-2	4	52.9
arena_text_factuality	1433	1442
Artificial Analysis Coding Index	44.2	48.7
browsecomp	51.4	65.8
BrowseComp (context management)	67.6	70
browsecomp_with_context_manager	67.6	65.8
browsecomp_zh	65	76.1
coding_arena_elo	1368	1418
critpt	2.9	11.6
deepsearchqa_f1	60.9	71.3
frontiermath_tier_4	2.1	18.8
gaia_level1	90.3	92.5
gaia_level2	73.6	74.8
gaia_level3	61.2	57.1
gaia_overall	76.7	77.4
gdpval	18.8	48.3
GPQA (unspecified)	79.9	92.4
GPQA Diamond	84	92.4
HLE	40.8	45.5
HLE (with tools)	40.8	45.5
HMMT 2025	90.2	99.4
hmmt_feb_2025	92.5	100
hmmt_feb_2026	84.1	97
hmmt_nov_2025	90.2	99.2
IFBench	61	75.4
imo_answer_bench	78.3	86.3
LiveCodeBench	86	89
longbench_v2	59.8	54.5
MCP Atlas	62.2	68
MMLU-Pro	86	87
OmniScience Accuracy	33	44.3
OmniScience Non-Hallucination	17.3	38.4
PaperBench	47.1	63.7
scicode	39	52.1
Seal-0	49.5	45
simpleqa_verified	27.5	38.9
swe_bench_bash	60	72.8
SWE-bench Multilingual	70.2	72
SWE-bench Pro	15.6	55.6
SWE-bench Verified	73.1	80
SWE-Perf	0.9	3.6
SWT-bench	62	80.7
TauBench V3 - Banking	18.8	11.1
Terminal-Bench 2.0	46.4	54
Terminal-Bench Hard	35.6	62.2
tool_decathlon	35.2	46.3
toolathlon	35.2	46.3
vectara_answer_rate	92.6	100
vectara_avg_summary_length	62	186.3
vectara_factual_consistency	93.7	89.2
vectara_hallucination_rate ↓	6.3	10.8
τ²-Bench	91	85.5
τ²-Bench Telecom (AA run)	90.6	98.7
τ³-Bench	69.2	12

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.