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

Across 51 shared benchmarks tracked on Vector Wire, Gemini 3 Flash Preview leads on 35 and Kimi K2.5 leads on 15 — Gemini 3 Flash Preview 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 Flash Preview vs Kimi K2.5

GooglevsMoonshot51 shared benchmarks35–15 head-to-head

Benchmark	Gemini 3 Flash Preview	Kimi K2.5
AA Agentic Index	49.7	52.8
AA Intelligence	38.7	36
AA-LCR	73	73
AA-Omniscience	10.1	-7.3
AIME 2025	99.7	96.1
aime_2026	96.7	95.8
apexAgents	27.7	11.5
arc_agi_1	84.7	65.3
ARC-AGI-2	33.6	11.8
arena_elo	1473	1450
arena_text_factuality	1452	1445
arena_vision	1260	1247
Artificial Analysis Coding Index	42.6	46.8
charxiv_rq	80.3	78.7
coding_arena_elo	1387	1436
critpt	8.6	3.1
frontiermath_tier_4	4.2	4.2
gdpval	35.2	38.3
GPQA Diamond	90.4	87.9
HLE	43.5	50.2
hmmt_feb_2025	97.5	95.4
hmmt_feb_2026	89.4	87.1
hmmt_nov_2025	93.3	91.1
IFBench	78	70.2
livebench	72.4	69.1
matharena_visual_math_overall	85.8	80.6
MCP Atlas	62	64
MMMU-Pro	81.2	78.5
OmniDocBench 1.5	12.1	88.8
OmniScience Accuracy	53.4	35.2
OmniScience Non-Hallucination	7.6	50
OSWorld-Verified	65.1	63.3
scicode	50.6	49
simplebench	61.1	46.8
simpleqa_verified	67.4	36.9
swe_bench_bash	75.8	70.8
SWE-bench Multilingual	72.7	73
SWE-bench Pro	49.6	53.8
SWE-bench Verified	78	76.8
TauBench V3 - Banking	20.8	14.2
Terminal-Bench 2.0	47.6	50.8
Terminal-Bench 2.1	58	45.7
Terminal-Bench Hard	38.6	34.8
toolathlon	49.4	27.8
vectara_answer_rate	99.8	92.2
vectara_avg_summary_length	90.2	112
vectara_factual_consistency	86.5	85.8
vectara_hallucination_rate ↓	13.5	14.2
VideoMMMU	86.9	86.6
τ²-Bench Telecom (AA run)	80.4	95.9
τ³-Bench	17.5	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.