Which is better, Gemini 3 Pro or Kimi K2.5?

Across 83 shared benchmarks tracked on Vector Wire, Gemini 3 Pro leads on 52 and Kimi K2.5 leads on 30 — Gemini 3 Pro leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Gemini 3 Pro vs Kimi K2.5

GooglevsMoonshot83 shared benchmarks52–30 head-to-head

Benchmark	Gemini 3 Pro	Kimi K2.5
AA Agentic Index	52	52.8
AA Intelligence	40.6	36
AA-LCR	73	73
AA-Omniscience	15.3	-7.3
AIME 2025	100	96.1
AIME 2026 I	90.6	92.5
aime_2026	91.7	95.8
apexAgents	18.4	11.5
arc_agi_1	75	65.3
ARC-AGI-2	54	11.8
arena_elo	1485	1450
arena_text_factuality	1481	1445
arena_vision	1290	1247
Artificial Analysis Coding Index	46.5	46.8
browsecomp	59.2	74.9
BrowseComp (context management)	67.6	74.9
browsecomp_with_context_manager	59.2	74.9
browsecomp_zh	66.8	62.3
charxiv_rq	81.4	78.7
coding_arena_elo	1438	1436
critpt	9.1	3.1
cybergym	39.9	41.3
deepsearchqa_f1	63.2	89
FinSearchCompT2&T3	49.9	67.8
Fortress	41.7	41.1
frontiermath_tier_4	18.8	4.2
gdpval	34.2	38.3
GPQA Diamond	91.9	87.9
HLE	45.8	50.2
HLE (with tools)	45.8	51.8
hmmt_feb_2025	97.5	95.4
hmmt_feb_2026	86.4	87.1
hmmt_nov_2025	93.3	91.1
IFBench	70.4	70.2
imo_answer_bench	83.3	81.8
InfoVQA (val)	57.2	92.6
livebench	73.4	69.1
LiveCodeBench v6	90.7	85
longbench_v2	68.2	61
LongVideoBench	77.7	79.8
LVBench	73.5	75.9
matharena_visual_math_overall	84.2	80.6
mathvision	86.1	84.2
MathVista	89.8	90.1
MCP Atlas	70.3	64
MMLU-Pro	90.1	87.1
MMMU-Pro	81	78.5
MMVU	78.9	80.4
MotionBench	70.3	70.4
multichallenge	65.7	61.4
MultiNRC	59	35.2
OCRBench	90.3	92.3
OJ-Bench (cpp)	68.5	57.4
OJBench (cpp)	68.5	57.4
OmniDocBench 1.5	88.5	88.8
OmniScience Accuracy	55.8	35.2
OmniScience Non-Hallucination	10	50
scicode	56.1	49
SEAL VISTA	51.5	41.9
Seal-0	45.5	57.4
simplebench	76.4	46.8
simpleqa_verified	72.9	36.9
SimpleVQA	69.7	71.2
swe_bench_bash	74.2	70.8
SWE-bench Multilingual	68.7	73
SWE-bench Pro	43.3	53.8
SWE-bench Verified	78	76.8
Terminal-Bench 2.0	54.2	50.8
Terminal-Bench Hard	56.9	34.8
tool_decathlon	36.4	27.8
toolathlon	36.4	27.8
vectara_answer_rate	99.4	92.2
vectara_avg_summary_length	101.9	112
vectara_factual_consistency	86.4	85.8
vectara_hallucination_rate ↓	13.6	14.2
Video-MME	88.4	87.4
VideoMMMU	87.6	86.6
WideSearch (item-f1)	57	72.7
WorldVQA	47.4	46.3
ZeroBench	8	9
ZeroBench (w/ tools)	12	11
τ²-Bench	98	85.4
τ²-Bench Telecom (AA run)	98	95.9

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.