Which is better, Gemini 3.1 Pro or Kimi K2.5?

Across 79 shared benchmarks tracked on Vector Wire, Gemini 3.1 Pro leads on 68 and Kimi K2.5 leads on 11 — Gemini 3.1 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 3.1 Pro vs Kimi K2.5

GooglevsMoonshot79 shared benchmarks68–11 head-to-head

Benchmark	Gemini 3.1 Pro	Kimi K2.5
AA Agentic Index	23	52.8
AA Intelligence	47.7	36
AA-LCR	79	73
AA-Omniscience	32.9	-7.3
aime_2026	98.3	95.8
apexAgents	33.5	11.5
arc_agi_1	98	65.3
ARC-AGI-2	77.1	11.8
arena_elo	1486	1450
arena_text_factuality	1471	1445
arena_vision	1280	1247
Artificial Analysis Coding Index	68.8	46.8
baby_vision	51.6	36.5
baby_vision_with_python	68.3	40.5
browsecomp	85.9	74.9
BrowseComp (Agent Swarm)	85.9	78.4
BrowseComp (w/ Ctx)	85.9	74.9
browsecomp_with_context_manager	85.9	74.9
charxiv_rq	89.9	78.7
charxiv_rq_with_python	89.9	78.7
coding_arena_elo	1447	1436
critpt	17.7	3.1
cybergym	38.8	41.3
deepsearchqa_accuracy	60.2	77.1
deepsearchqa_f1	81.9	89
Fortress	29.8	41.1
FORTRESS (Adversarial)	65.2	54.1
FORTRESS (Benign)	98	98.3
frontiermath_tier_4	16.7	4.2
gdpval	23.3	38.3
GDPval-AA v2	962	1009
Global-MMLU-Lite	92.7	84
GPQA Diamond	94.3	87.9
HLE	51.4	50.2
HLE (with tools)	51.6	51.8
hmmt_feb_2026	94.7	87.1
hmmt_nov_2025	94.8	91.1
IFBench	77.1	70.2
imo_answer_bench	91	81.8
livebench	79.9	69.1
LiveCodeBench v6	91.7	85
matharena_visual_math_overall	89.4	80.6
mathvision	89.8	84.2
mathvision_with_python	95.7	85
MCP Atlas	78.2	64
mcpmark	55.9	29.5
MMLU-Pro	91	87.1
mmmu_pro_with_python	85.3	77.7
MMMU-Pro	83	78.5
multichallenge	71.4	61.4
MultiNRC	64.7	35.2
nl2repo	33.4	32
OJBench (python)	70.7	54.7
OmniScience Accuracy	55.3	35.2
OmniScience Non-Hallucination	50.1	50
OSWorld-Verified	76.2	63.3
scicode	59	49
simplebench	79.6	46.8
simpleqa_verified	77.3	36.9
strongreject	98	99.5
SWE-bench Multilingual	76.9	73
SWE-bench Pro	54.2	53.8
SWE-bench Verified	80.6	76.8
SWEBench Pro (Public)	54.2	50.7
Tau 3 Banking	16.5	14.2
TauBench V3 - Banking	21.4	14.2
Terminal Bench 2.1 (Best Harness)	73.8	51.3
Terminal-Bench 2.0	68.5	50.8
Terminal-Bench 2.1	74	45.7
Terminal-Bench Hard	68.5	34.8
tool_decathlon	48.8	27.8
toolathlon	48.8	27.8
V* (w/ python)	96.9	86.9
vectara_answer_rate	99.4	92.2
vectara_avg_summary_length	107.7	112
vectara_factual_consistency	89.6	85.8
vectara_hallucination_rate ↓	10.4	14.2
τ²-Bench Telecom (AA run)	99.3	95.9
τ³-Bench	67.1	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.