Which is better, DeepSeek-V4-Pro or GLM-5.2 Full Open Source?

Across 58 shared benchmarks tracked on Vector Wire, DeepSeek-V4-Pro leads on 17 and GLM-5.2 Full Open Source leads on 40 — GLM-5.2 Full Open Source leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V4-Pro vs GLM-5.2 Full Open Source

DeepSeekvsZ.ai58 shared benchmarks17–40 head-to-head

Benchmark	DeepSeek-V4-Pro	GLM-5.2 Full Open Source
AA Agentic Index	63.3	45.7
AA Intelligence	53	52.6
AA-LCR	70	76.7
AA-Omniscience	-10.6	4.4
Agents' Last Exam	16.5	23.8
aime_2026	96.7	99.2
apexAgents	24.3	33.7
arena_elo	1457	1471
arena_text_factuality	1450	1460
Artificial Analysis Coding Index	59.4	68.8
AutomationBench (Public)	12.8	12.9
AutomationBench Public	12.8	12.9
coding_arena_elo	1582	1582
critpt	13	20.9
DeepSWE	12.8	46.2
DSBench-FullStack	41.8	61.8
DSBench-Hard	31.1	54.5
FORTRESS (Adversarial)	36	71.3
FORTRESS (Benign)	98.5	90
gdpval	49	50.3
GDPval-AA v2	1307	1514
Global-MMLU-Lite	89.3	89.2
GPQA Diamond	90.5	91.2
HLE	48.2	54.7
HLE (with tools)	48.2	54.7
HLE (wo / w tools)	48.2	54.7
hmmt_feb_2026	95.2	92.5
hmmt_nov_2025	94.4	94.4
IFBench	76.5	73.3
imo_answer_bench	89.8	91
itbenchSre	38.3	42.7
livebench_agentic_coding	42.6	51.8
livebench_coding	70	79.7
livebench_data_analysis	74.5	73.7
livebench_instruction_following	62.4	62.3
livebench_language	78.1	76.2
livebench_math	90.7	89.8
livebench_reasoning	82.7	78.6
MCP Atlas	74.2	82.6
nl2repo	38.5	48.9
OmniScience Accuracy	43	24.3
OmniScience Non-Hallucination	12.2	73.7
Program Bench	47.8	63.7
scicode	50	50.5
simplebench	50.9	58.8
simpleqa_verified	57.9	38.1
strongreject	98.6	98.5
SWE-bench Pro	55.4	62.1
SWEBench Pro (Public)	55.4	62.1
Tau 3 Banking	26	26.8
TauBench V3 - Banking	30.1	34.6
Terminal Bench 2.1 (Best Harness)	64	82.7
Terminal-Bench 2.1	72.1	82.7
Terminal-Bench Hard	46.2	50.8
tool_decathlon	52.8	48.2
toolathlon	51.8	48.2
τ²-Bench Telecom (AA run)	96.2	99.1
τ³-Bench	25.8	26.8

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.