Which is better, DeepSeek-V4-Pro or Kimi K2.5?

Across 54 shared benchmarks tracked on Vector Wire, DeepSeek-V4-Pro leads on 46 and Kimi K2.5 leads on 8 — 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-V4-Pro vs Kimi K2.5

DeepSeekvsMoonshot54 shared benchmarks46–8 head-to-head

Benchmark	DeepSeek-V4-Pro	Kimi K2.5
AA Agentic Index	63.3	52.8
AA Intelligence	53	36
AA-LCR	70	73
AA-Omniscience	-10.6	-7.3
aime_2026	96.7	95.8
apexAgents	24.3	11.5
arena_elo	1457	1450
arena_text_factuality	1450	1445
Artificial Analysis Coding Index	59.4	46.8
browsecomp	83.4	74.9
BrowseComp (w/ Ctx)	83.4	74.9
coding_arena_elo	1582	1436
critpt	13	3.1
cybergym	52.7	41.3
FORTRESS (Adversarial)	36	54.1
FORTRESS (Benign)	98.5	98.3
gdpval	49	38.3
GDPval-AA v2	1307	1009
Global-MMLU-Lite	89.3	84
GPQA Diamond	90.5	87.9
HLE	48.2	50.2
HLE (with tools)	48.2	51.8
hmmt_feb_2026	95.2	87.1
hmmt_nov_2025	94.4	91.1
IFBench	76.5	70.2
imo_answer_bench	89.8	81.8
livebench	73.6	69.1
MCP Atlas	74.2	64
MMLU-Pro	87.5	87.1
nl2repo	38.5	32
OmniScience Accuracy	43	35.2
OmniScience Non-Hallucination	12.2	50
scicode	50	49
simplebench	50.9	46.8
simpleqa_verified	57.9	36.9
strongreject	98.6	99.5
SWE-bench Multilingual	76.2	73
SWE-bench Pro	55.4	53.8
SWE-bench Verified	80.6	76.8
SWEBench Pro (Public)	55.4	50.7
Tau 3 Banking	26	14.2
TauBench V3 - Banking	30.1	14.2
Terminal Bench 2.1 (Best Harness)	64	51.3
Terminal-Bench 2.0	67.9	50.8
Terminal-Bench 2.1	72.1	45.7
Terminal-Bench Hard	46.2	34.8
tool_decathlon	52.8	27.8
toolathlon	51.8	27.8
vectara_answer_rate	97.2	92.2
vectara_avg_summary_length	153.8	112
vectara_factual_consistency	91.4	85.8
vectara_hallucination_rate ↓	8.6	14.2
τ²-Bench Telecom (AA run)	96.2	95.9
τ³-Bench	25.8	66

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.